Master OpenClaw in 10 Hours [I Created 5 AI Employees]

This is the most comprehensive OpenClaw masterclass on YouTube. I say that with confidence because I've searched. No one has gone this deep. Every OpenClaw tutorial that I've seen maybe cover a setup, installation, how to use chat. Maybe they will teach you how to connect the tool. And some people just dump use case after use case on you for you to set up. That's fine if you want to run a chatbot. But if you want AI employees that can run your outreach, take your phone call, prepare your meetings, close your deals while you sleep, this is the course for that.

For those of you who don't know me, my name is Mani Kanasani. And I run a 7-figure AI agency in Vancouver, BC. I have a master's degree in cyber security and I was a CTO of a Web3 company. In the past three years, we have worked with businesses from teenagers running e-commerce stores all the way till Fortune 500 companies. So we know a thing or two about deploying AI agents in production and what happens when we don't take security seriously.

And I have put in a ton of time to make this as practical as possible and I've laid it out so you can jump to any section. You don't have to watch this in order. So we have about 23 chapters covered in this course. Let's take a quick look.

Chapter one is installation and setup. This is the fastest setup you'll find, plus how to get up to $1,500 in free API credits across AWS, Google Cloud, and Azure OpenAI.

Chapter 2, sub-agent configuration. How to set up your main brain model and cheaper muzzle agents so they work together without breaking your configuration. Trust me, OpenClaw breaks when you do this wrong. I'll show you the right way.

Chapter 3 is token optimization. The 8-layer stack that can take your monthly AI spend from $150 down to $10. Killing thinking mode, capping context, model routing, session discipline, prompt caching, heartbeat to Olama, sub-agent isolation. Each one stacks on the last.

Chapter 4, Security. And the one everyone's skipping. 17,000 exposed OpenClaw instances, two unpatched zero days. And I have a master's in cyber security, so I'm giving you a hardening guide that stops 80% of the attacks in 15 minutes. Mac Mini guide, VPS guide, everything.

Chapter 5, Business Brain Level 1. This is where you introduce yourself to your AI. If you skip this, your agent doesn't know who you are, what to do, or how you work. We fix that in 5 minutes.

Chapter 6, Business Brain Level 2, strategic context, your mission, your products, your team, your business model, your communication style. 15 sections of professional context so your agent can actually make business decisions.

Chapter 7. Business Brain Level 3. This is the enterprise grade. Operational playbooks, decision trees, escalation paths, historical meetings, competitive landscapes. This is what turns your agent from an assistant to autonomous operator.

Chapter 8. Memory Architecture. Soul, identity, user context, agent, tools, daily memory logs. I'll walk you through every file, what goes where, how to prevent memory decay because OpenClaw will quietly drop your most important instructions if you don't set this up right.

Chapter 9. Mission Control Dashboard. We build a real command center from scratch, so you always know what your agents are doing. Full UI, agent profiles, live status monitoring.

Chapter 10. Integrations. GitHub, browser control, email via agent mail. Wiring everything together so your agents can actually interact with the real world.

Chapter 11. Task Management. A full Kanban board where your agents pick up tasks, move them through stages, and file reports. To do, doing, needs input, done. All autonomous.

Chapter 12. Builder, Orchestrator, and Executor. This is the framework that changed everything for me. Your builder constructs platforms and your orchestrator, that's OpenTlo, manages workflows. And your executors are specialized agents that do actual work. Once you understand this separation, everything clicks.

And Chapter 13. Business Operation Loop. Seven stages. Outreach, discovery, proposal, sales, onboarding, retention, and competitive intelligence. Every AI employee we build maps directly to one of these stages. You are not building cool tools. You are building an autonomous business.

Chapter 14, Meeting Intelligence. We built Jane, a meeting assistant, with access to over 1,650 of my real transcripts. Every action item, every follow-up, every insight from every meeting you ever had accessible to your AI. Proposals, prototypes, call prep. All generated from your meeting data. Autonomously. We will build that.

Chapter 15, AISDR and Email Engine. We build an AISDR that scrapes leads from natural language. Just say, law firms in Vancouver, accountants in Vancouver, and it goes out, finds them on Appify, gets their email, gets their phone number. Personalizes every single word, and tracks every open, every click, every reply, and every bounce. You only get involved when someone books an appointment.

Chapter 16. Voice AI and Phone Agent. We build Alexa, a phone agent that makes and receives phone calls. Inbound, outbound, speed to lead. As soon as someone fills a form, it calls them. And the analytics dashboard is amazing. It's better than any voice AI SaaS could give you. I'll show you the whole build.

Chapter 17, Typeform Clone and Webhooks. Right in the middle of all this, we build a Typeform clone and replace what most you'd use Make.com for. Webhooks and functions, custom endpoints, all inside OpenTlo, built in minutes.

Chapter 18, Multi-Agent Orchestration. Two agent systems dispatching work to each other. Your OpenClaw agent assigns tasks to CloudCode agent. And CloudCode agent will assign tasks to OpenClaw agent. We see our application Superbase assign tasks. And these agents build it, ship it back, and OpenClaw will review it. If it's not right, it sends it back with nodes. Fully autonomous.

And Chapter 19 is agent-to-agent communication. Discord as your AI team's Slack handle. I'll show you how to set it up. A server, I will meet my agent, set it up. And these agents talk to each other. They send alerts, request, send updates, and escalate when they need you. Every agent has its own bot, its own channel, its own presence.

In Chapter 20, I discuss how these agents can talk to each other. Agent communication. Instead of you going back and forth, agents asking, "Is this okay?", you let them sit in a council and discuss it among themselves. They debate, they decide, they come to you only when they need approval.

And Chapter 21, Command Center Pipeline, an eight-phase autonomous pipeline that builds, deploys, and reports while you sleep. This is the full orchestration engine running 24-7.

And Chapter 22, Cronjobs and Automations. Polling, webhooks, heartbeats, scheduled tasks, Telegram notifications. We wire everything so your agents stay alive and responsive around the clock.

And finally, Chapter 23. Claw Alley. I built a platform where AI agents transact with other AI agents in real time. Real crypto, real blockchain, real money from my wallet. USDC on Base network. I will show you a live transaction. One agent buying a service from another AI agent. And payment was confirmed on-chain. I believe this is the future of agent-to-agent commerce.

And if you want the free guide that covers everything we discussed today, you can find that and link in the description below. Now, let's start building.

So what is OpenClaw? OpenClaw is an open-source AI assistant platform that is one of the fastest-strong repos on GitHub. Help it has more than 265,000 stars at the time of this recording and all the dogs, all the releases. If you look at the past two months, they've released more than 59 things. And if you open it, if you look at the updates, new models being supported, new tools for PDF analysis, Telegram voice management gating. So every day, almost sometimes twice a day, they're releasing new things. So they're fixing security issues. There are a lot of things that are going on right now. And do we discuss any of that? No, not really. What we are going to discuss is they cover all the installation, APIs, models, configuration, security, and everything. But this master class skips all of that. Of course, we'll talk about, briefly, about installation and security, which is a big thing. Instead, it shows you how to wire OpenClaw into real business operations, outreach agents, meeting prep, competitor intelligence, autonomous builds, and retention workflows. You will see flows after a discovery call happened. It will go out and create a prototype for you, create a proposal that you can send out to a client, all without you lifting a finger. And we will build AISDR where you can, in natural language, explain who you want to target. And it will go out, scrape the list, personalize emails, and follow up with them 5, 6, 10 times, however many times you set. And only get you involved when they book an appointment. How amazing is that?

So everyone is talking about how to play chess, how to ask for weather. We are building business operation intelligence. We are building business operation system. We are building business operations loop with seven business operations completely done by AI agents.

So what's covered in this video? We are going to discuss 12 sections in detail. And the way we arrange this video, as you can see, this is the most comprehensive video that you can find on OpenTlo right now. And every section is designed in such a way you don't have to watch it in sequence. And if you are more interested in 11th or 12th section, you can watch that or 5th or 6th section, you can do that. They are not really interlinked or connected to them in any way. But that said, first two sections will cover installation and security, which most people are conveniently ignoring. And we will have agent foundations and dashboard in three and four. What is foundation? We will talk about memory, self-evaluation, and real-time machine control dashboard for your AI workforce. And what do sections five through seven cover? We will talk about how to wire these tools together, you know, build a task board or map every feature that we have and every employee that we are going to build to a seven-stage business operations loop. That's what I was saying in previous section. Related to outreach, discovery, proposal, sales, onboarding, upselling, retention, competitor intel, all of that we will map to a business process. And AI employees, this is the meat of the. We will talk about meeting assistant, phone assistant, email assistant, AISDR, and we will build a full-time autonomous workers, 24-7 working FOI and multi-agent and command center. This is something really cool. Two AI systems dispatching work to each other and an eight-phase autonomous pipeline that builds while you sleep. And this works with CloudCode as well because Anthropic came out and said you cannot use a Roth token, meaning your Cloud Max plan with OpenFlow. But we figured out a way where OpenFlow can assign work to CloudCode agent, and CloudCode agent has a way to work on the project, dispatch it, you know, dispatch, claim the dispatch that's done by OpenFlow agent, work on it, give back, and OpenFlow agent can review or you can review yourself. And these are the changes that I need. And it can send back the work saying this is not working, or you can say it's not working and it will rework. How cool is that? And it's possible to do that today. And finally, Alchemist and Automations. You already have cron jobs. Everyone is talking about cron jobs. That's why I actually pushed it towards the end. But this is one of the sections that I really love. Alchemist is replace a SaaS subscription with custom AI modules. And this is something I found very little talked about on YouTube. That's why even though it's a focus, I'm only going to cover one or two simple tools in this video. This deserves a standalone topic because I uploaded my six-month statements, bank statements, credit card statements, and OpenClaw came up with all the things that I've been paying. GoHighLevel for $500, and it looked at a fiduciary school platform for $300, and then came up with, "Okay, you are paying this platform, this platform, this platform, around more than $1,000, but we can build all of that today." And it blew my mind. That's $1,000 of saving every single month. Can you even imagine what this can do if you give it one month's time? And that's one of the reasons I've been so focused on building and building. And I now learned a lot about both OpenFlow and CloudCode. This is the first masterclass tutorial that I'm doing.

For those of you who don't know me, my name is Mani Kanesani, and I'm from Vancouver, BC. I run an AI agency and I've been documenting my journey on YouTube for the past three years. And in this past three years, we have served a ton of companies, starting from teenagers running e-commerce stores all the way till Fortune 500 companies and a lot of businesses in between. So we know a thing or two about running, deploying AI agents, and we do take security seriously. I have a master's degree in cybersecurity and I worked as a CTO of a Web3 company before I started my AI agency. And if you want to get all of this without building, with just one click of a button, and if you want to vote the entire system, a better system, you can join my school community that I'm launching, Agents in a Box. And without further ado, let's start building.

Step one is installation. And I think there are around thousand videos at least people talking about how easy this is to install. And they're not wrong. You know, I even I made a setup video. But what's different about my setup video is I show you a way to get $1,500 free credits. If you haven't watched that, I don't want to repeat the whole thing. You can watch that video. You will get $200 in Amazon AWS credit, which you can use for VPS. And you will get $300 for Google Cloud about credits, which you can use for AI models like Gemini 3.1 or, you know, flash models if you want to keep this for longer. But all of my viewers who watched that video kind of complained, Google has been rate-limiting. Even though they give you $300 credit, there are limits to what you can realistically use for OpenTlo. And I already showed you one more way to get $1,000 credit from Azure OpenAI. Now, you cannot use that as a base model for your OpenTlo, at least last I checked, it's not possible. But if in future, I'm sure there will be support for that. You've seen how many, you know, at the pace that they're doing these releases. But I figured out a way to use that as a sub-agent. We can use the same $1,000 credits that you get, and covered in the video, as a sub-agent, where we can use whatever main agent you're using. I'm using my GPT 5.3 Codex, the ChatGPT $200 plan, for my main agent. But if I want to use OpenAI for a good OpenAI model like 5.2 or something for my proposal generation or analysis or research tasks, because I have $1,000, I can use it as a sub-agent. And we will talk about what these main agents, sub-agent, the kind of pros and cons of using all of this. So I made another video about sub-agent configuration and how I'm using. You don't have to watch this. I will cover what I've done. In here. So installation, I just want to quickly touch upon how easy this is. Whether you were using Mac Mini or MacBook, old MacBook, laptop lying around, or Windows, whatever you are looking for, just change it and it will give you a one-liner. And it is so powerful because first time when I was using this, I have two Mac Minis. The first Mac Mini, I was, I didn't know it was installs everything like a one-liner thing. I was installing my Homebrew, I'm installing my NPM and all of that, and it took a long time. But the next time, the second time I got it, I just copied this command and pasted it and hit run, and it did amazingly well. So you don't have to. You just, if you are using VPS or if you are using, you can watch my other video. But it's as easy as it's on copy and headset. And they also recently released a companion app, which is in beta, which will, once you download it, will stay on top like a menu bar for header. It will be there. You can click it. It crashed before I showed you. I wanted to record it. There's a fun little app. And today is only March 5th, so I'm sure in coming days they will release an actual version, not a beta. So we will make a video when it comes out. But really quick, before you install, you need to decide what this thing is going to do. And this is the question I get a lot. Mac Mini or VPS? Or when I say Mac Mini, you can use any MacBook that you are not using, or a notebook, or a Windows laptop. It doesn't matter. Any standalone machine. Or is it VPS? If you take side-by-side comparison, Mac Mini, the Costco Mac Mini will be, I work at one time, $600 to $800. VPS is $5 to $50. And the speed here is instant. You get instant local access. But VPS will have some latency. Some of this is not even noticeable. Data privacy, everything is in your hardware, feeble, Mac Mini. But VPS, it lives in software and system. And it is always on. Mac Mini, you need to have it. You need to change the energy settings and you need to keep it all powered. VPS is 24/7 by default. And if you want remote access, you need DataScale or other applications on your iPad or iPhone to monitor what's going on and not be easier right here. But VPS is, you can access such software anyway. And what is this best for? In the end of the day, one of the things that I haven't mentioned is something they released two days ago, browser control. I'd love browser control after the update. We have building some amazing applications, use cases. You will see that in chapter 3D4. And what is this best for? I would say Galway control, privacy first, in tool control. And VPS is no hardware and it's a quick start, low barrier to entry. You can get started for free in some of the platforms if you want. And here's what I recommend. If you are serious about this, if you are building AI employees that handle your email, full calls, business data, and you want browser control features, everything, you are, I'm sure here, and you want to stand it up, get a Mac Mini. That doesn't mean you know, you have to get it. And it's not necessarily have to be you. I am not too many rounds open from beautiful or any laptop that's lying around in your home. And you can find them for six, three, four hundred as well. Your data never leaves your house. You'll be hungry. And once it's set up, it's set up. If you are just getting started, testing things out, and you don't want to buy high-rated, a VPS works perfectly fine. $5 to $25 a month on DistroLotion gets your machine that runs OpenDlo 24/7. And I need a video. If you just scroll down, it is $20 a few days. You get $200 in the US credit. So you can use that to get started. So a VPS works perfectly fine and you can SSH from anywhere and also prancio open 24/7. That only trade-off is your data lives on someone else's server. Migration is a bit headache. And when you, when we get to chapter 4, you'll see why this matters. My setup, I used to map minis, one for production, more for testing. But I started on a BPS. So don't let hardware stop you. Just pick one and keep going. Everything in this course works on both, except the final chapter and some browser control tense. I will slide it when it max specific or BPS specific. But let's keep going. That's it for installation. I know it's one of the fastest things. But it doesn't matter if you are using VPS or web UI, whether that's a VPS or Mac Mini, be everything that we are covering today, you will be able to take full advantage of this. I will not show both of them. But both of them can be configured on Telegram just as easily. Right? So we will see how to. Essentially, most of the time, the things that I'm saying, even if I use web UI, you can see the same things in Telegram as well. And we are going to build an amazing application. And that works completely fine even if you are using a web-based, even if you are using a VPS and you are only using terminal. Do you want?

So, one thing that I wanted to follow before we go to the next sections, if you don't watch this video, you will miss this part, which is really, really important. OpenClaw has a habit of breaking when you ask it to make config changes. So, I'm not sure if you face this. Anytime I ask it to, "Oh, can you add this sub-agent? Can you add this model?" I really love Alex Finn's content, but his best advice is, "Oh, ask your OpenClaw agent, it will know what to do. Tell your OpenClaw agent, do this, it will do that." Maybe his Opus 4.6 did, but even I'm using Opus 4.6 for a while now, and even that was breaking its own configuration. The issue being, we are using brief search or Perplexity search at best, and the results that would come up is some outdated ones from the time they it's called flawed bot, right? Two months ago, and it will try to use that info and try and change the configuration. A bunch of things changed. It's like I'm giving you a basic example. Might be one of the reasons it might be not. It's like garbage in, garbage out. If the information itself is wrong, no matter how powerful of a model you're using, it will break. I came up with this prompt: "Define agent and agent are less minimal config only." So what you would do is, you will copy the whole thing, you will copy the prompt, and don't ask OpenClaw to figure it out. Wherever you have your OpenClaw agent, you go to your agent and say, "I want you to create a sub-agent, and this is the guideline." And you paste the entire prompt that is left. This is the guideline, step one. And it will give this output: "I set this up for you. Applied config changes. Researcher kept research at minimum. Then added model, sober write, researcher sub-agents model." Now it has other agents that we ask, "Do you have the API key?" No, it won't. We need to set it up. So, the reason I can stop by, "Okay, tell me the best way to and secure way to give you API keys." So I will show you the secure way to give API keys. And I want you to realize this is the keys that I'm using. If by any chance if I don't blur it, the chances are I will hook the key as soon as the video is done. So set locally, this is practical and most safe, right? And the way we do that is we need to first export it in our terminal. And so, for example, I want to give it Anthropic key. Okay. And where is my Anthropic key? I can go to, let's say, we want to give the Anthropic API key. And what would I do to give this an Anthropic API key is I will just do an export Anthropic API key, something like this in my terminal. And then I need to restart the gateway and say, "Hey, right now OpenClaw key, Anthropic API key is available as an environment variable. I want you to use that and choose that for sub-agent." And OpenClaw is the key to work with. I will show what I mean by that. So I went to Anthropic and I created a test key, which I am going to delete. And also this amount is just for tutorials, and I will delete it once we are done anyway. So you can see the number of credits we have, it's around $8, nothing much. So I can come back and I can go to the terminal. I was running into some issues. So as you can see here, I tried an approach, I didn't work. And then I did this echo export this and ZSHRC. You can take a screenshot or to get this. And then source ZSHRC and then I restarted the gateway, force I do restart. And then I said, "Use Sonnet as sub-agent to rest, right essay." I asked you on it. I dispatched your essay request to researcher sub-agent. And it's an OpenFlow is a kind of tool. I'm not exactly sure if this world. We can go to our Cloud API and check. Go back to the link. Even 8.67 never used. So this is the thing with OpenClaw. It said, "Oh, I will use it." And I used it. So I will say, "You have to use sub-agent Sonnet 4 or newer." This is how to use that messages API. And look for Sonnet 4.6. Or if you use Opus 4.6, I'm okay with that too. And the thing is, you need to also give the documentation. So send it. So it restarted the gateway. So I want you to create an essay why OpenFlow is amazing. And so it finally worked. I perfect pipe kick to e-search as a patient. So I did not know why it didn't work previously. We did all the right things. Did you actually use this? And I gave it the entire thing again. And then I said, "Yes, so can you use Sonnet, OpenNoise leverage, you've been missing?" So when I go back, my even gone down a little bit. So you can see. And we can also see it used Cloud Sonnet 4.6, like here. So it is a bit tricky. That's why I had to cover it even before everything else. So everyone knows it's not very easy to add sub-agents and spawn sub-agents.

Let's go to the next topic. Now, that's only one way of saving tokens, having this brain, this expensive model, so to speak, or your default model, which is really good and better than your other models as a brain. And models that are not as expensive, like if you want to use a Minimax or a Kimiki 2.5, these can be your muscle agents who can do the work, right? So this setup has helped me tremendously when I was starting out and I was hitting limits with my even with my Cloud Max $200 plan, I was hitting in limits. So I started doing Opus as a control and all the agents as the muscle. So that is only one. But we have eight more eight-layer token optimization stack. Now, these are the things that I found out after researching a lot of videos. I looked at some of the well-performing videos. Of there are a lot of people who've done videos on token optimization. Am I using everything? No. Should you use everything? I don't think so. But it is there if you want it. I know people who, you know, spending $5 a day is expensive for them. And it is, if you ask me, it also depends on what value you are getting from OpenFlow, right? If you are not using it to the extent that I am, if you are not using it to the extent of my client Ryan, who made like in four days, he made close to $62,000 and um, in new client revenue and savings. So what I'm trying to say here is, there are places where it pays to go all out, you know, getting that $200 Cloud Max plan or $200 Codex plan like I'm doing. But there are times where you just use AI for the basic and the most. Where you are using AI not for income generating or revenue generating activities, but your content work or anything to do with productivity, you use those models. So what are these eight steps that can bring your $150 a month to $10 a month? I know that's extreme. That's a big promise. But I want to go through one by one. The first one is killing thinking mode. And I have details of everything here. Capping the context window, model routing, which we talked about briefly. Session reset discipline, lean session initiation, heartbeat to Olama, and prompt caching, and sub-agent isolation. So when you add one on top of each other, I also added, you know, how much savings each one brings you and the time it takes to do that. So when you kill thinking mode, you are saving 10 to 15 times per call and it takes you literally 30 seconds to set it up. So context, uh, you know, capping context window, saying this is only the set context that you would get, no more. And when you have that, you are saving 20 to 40% of your usual things. And you are saving, and you are doing that just with one configuration line. That's the same thing with the kill thinking mode as well. It's just one configuration line. And doesn't take longer than 30 minutes. Will you lose a lot? No, if you are doing something. I personally haven't done this, but I would do maybe three of these at least, right? So it's case by case basis. You have to decide whether this is right setup for you. So how do you kill thinking mode? Let's say you have a thinking model, Kimiki 2.5 thinking, right? And it, when you ask a question, it thinks and thinks. And anytime it is thinking a lot, it is actually wasting tokens, right? So if you, what if you look at this page, I have one line, "disable in config." And every layer or every thing that's mentioned in this section helps you. Every layer, every hack, everything mentioned in this document is detailed. So you can read: "Extended thinking is the single biggest hidden cost in OpenClaw. When enabled, every API call generates thousands of reasoning tokens before the actual response. Most routine tasks don't need it. So this one optimization alone can save more than all others combined. And if you do one thing from this guide, you do this." And how do you do that? You need to go to OpenClaw.json and you just need to update this. And if you are technical, you can pull this off easy. You know how to update config. But for those of you who are not very technical, how do you open this? You go to your terminal and type `open .openflow`. This is a hidden folder, remember? And when you click that, it will open it up. And if you are following everything I've been telling you, the way you export the API keys, setting them in the as an environment variable, and asking OpenFlow to find them, if you are following that method, this config file should not have any API keys. I do have some API keys because I, it's a Mac Mini and I was lazy. Okay, everyone. So what I'm telling you is, if you copy the whole thing and go to any agent, any chatbot frontier, your model, ChatGPT, Cloud, doesn't matter. You ask it to make sure you don't have the API keys, okay? You will have to rotate it. It's in chat. You don't want all of that. So only do that if you have not used it in the past. I used API keys in. The chat. If you pick your model section, you can copy that section and go paste in your chatbot and ask, "This is the guideline, how to disable thinking." Usually, it's one line about models, and you just need to add this here: `open thinking. column double quotes curly brace thinking i mean reasoning off i believe uh if you look at thinking type disabled close that's it`. This one line, followed by a comma, should be enough. So I will see if we can, I can show you in the actual. So I went to Claw and said, "This, I'm trying to save tokens and turn off thinking in my OpenClaw.json. How to do that based on your masterclass guide?" Add this to your OpenClaw.json. It says, "Before agents, defaults, model, this, and before the model starts, you add thinking and colon curly braces, type disabled." And that's layer 1 from 8-layer stack, the single biggest cost saver. And then layer 2, right there, just below that, you can set context tokens 50,000. And that is two things, just like that. So anytime you have a question, you only need to go to your chatbot, whichever you're using, and asking it to, "Hey, implement this guide. This is the URL." And context window gap, capping context window, model routing. And this is interesting. They added this recently in one of the updates. OpenClaw uses Sonnet for everything, right? Sonnet is excellent, but overkill for checking file status, running commands. A routing monitoring. Haiku handles these perfectly at 12x less cost. So instead of having a default Sonnet model, you can use Haiku model. I was happy 80% of the time. And if you watch my one of the two videos that I linked there, the second video of, I was using Sonnet and I was happy 95 or 100% of the time. When I switched to Haiku, it was super fast and it was really helpful too, and really cheap. And once you add these models, you need to have a rule, a routing rule, something like this: "Model selection rule: Default, always use Haiku. Switch to Sonnet only when needed. This, when in doubt, try Haiku first." And it's a lot of cost saving. And also, I wanted to include some free and ultra-cheap alternatives, which is Gemini Flash, task free tier available, you can use that for heartbeats, classification, simple routing. And Deep Seek v3, you can use that for code generation, reasoning at a fraction of cost. Llama 3.2 free, it's a local. Kimiki 2.5 is available on Nvidia API for free at the time of creating this. If it's changed, you still have these models. That's one of the reasons why I did not make an actual tutorial on this. You can find a bunch of videos on YouTube who talk about how to use Kimiki 2.5 for free. And Cloud Haiku is general default for 80% of the tasks that you should be using. So you can use Opus or Sonnet for initial strategy and route first draft to Kimiki and use Haiku for everything else. Three tiers, one configuration, you cannot go wrong when it comes to saving tokens.

Let's briefly talk about session reset discipline. You can check whether this session is bloated by `/status` and tells you it's a big session or a small one. And right now, you see the context is 71k and 26%. And the more I use it here, the more tokens have. Burn money. Why are you not practicing what you preach? You said 50k context window. Yes, I did. But I'm on a Codex GPT 5.4 and I have a like a pro plan, a $200 plan. Even if I wanted to, I'm hitting not hitting the limits. So it's okay for me. But not everyone is using this plan, right? So you need this more than me. And also, you know, check current usage, compact the session, keeps context and reduced token. And you can use `/compact` to compact the conversation. Basically, think of it like summarizing everything what was happened in the past 30 minutes or so. And the lean session initiation is the next strategy, which is default load behavior. It loads 50 KB of history on every message. You need to optimize to load only 8K of what actually matters. So add this rule on every session load only these files: user identity, memory of that day, if it exists. Do not auto-load long-term memory, session history, prior messages, previous tool outputs. Whenever user asks about prior context, then we can use memory search on demand. Pull only the relevant data snippet, don't load the whole file. And update daily memory logs at the end of session with what you worked on, decisions made, blockers, and next steps. This will save you a lot of tokens, a lot of, you know, it gives you a lot of peace, really, because it drives you crazy on how it used to be a really big problem before OpenClaw, right? Now, OpenClaw has a memory and it persists. But AI has this tendency to repeat its mistakes, right? Even CloudCode, to some of the smartest models, if you don't have that self-improvement, if you don't have that cognitive memory built in, it doesn't catch its errors and create a lessons learned like a gotcha. It won't remember. It will repeat the same thing. So update memory daily logs at least at the end of each session. So what was before all of this? 50 KB to 3 million wasted per session for at least 40 cents per session history bloat over time. But now 8KB only loads when needed, 5 cents per session. Lean daily memory locks and heartbeat. You can use Olama, you can use Gemini Flash, you can use really cheap model like Deep Seek v3 if you want to go crazy. But why pay for something you can have for free? Again, not doing it because I don't have the need to. So I added the install steps, how to pull a lightweight model just from your command line. And configuring heartbeat also, you go to OpenClaw.json and add this snippet. And you can verify `olama serve` and then you can verify it's working. If heartbeat was running every minute on paid API, that's 1,440 API calls per day. $5.15 just for "are you alive" checks, right? And it's not even that. Even if you use some of the cheapest models available, you would still be charged around $5 with just that one thing. And next says prompt caching. Prompt caching token cost and drop a KPI. So next we have prompt caching. And you can see your system prompt and reference materials get sent every API call. Prompt caching charges only 10% of reused context within a 5-minute window. So you need to set this up. And what to cache and what to not. System prompts, you can do all of that. Do not cache dynamic data, daily memory logs, recent user messages, and all of that. Why do we even do this? Because like I said, it is not as expensive as regular tokens. So if you look at the cache hits, refreshes, and also input tokens and output tokens. If you look at IQ, the cache is much lower. So if you go to Anthropic API, for an example, so and I've, if you read it here, Anthropic prompt caching allows input tokens are built when you reuse the same prompt prefix. Here's the token structure: input $3 per million tokens, outputs $15 per million tokens. But if you read here, normal token input 100% is $3. Cache right the first time. Cost versus normal. Because we are doing this, we are going to get a benefit. Towards the end, they charge you a bit of extra $3.75. Then when you reuse it, you're only charged 10%, like 30 cents. When you really load the whole thing over and over again in a one-hour conversation session that you have, if the caching is off, you are saving essentially saving at least for this part, the parts that are reloading 90% of the cost. It's really great. People are missing it. Even if you have a model that is not as expensive, you should still do this because even if you take Kimiki 2.5 API price in. And this is the final one. We have sub-agent isolation. What is this? Multi-agent coordination is expensive. It is almost 3.5x more tokens than single agent setups. And what is the fix for that? Fix is you isolate subtasks with `/spawn`. So each agent runs in clean context. So instead of asking your main agent to do everything, if you look at the example that we were doing earlier, I was asking it to use Claude Sonnet to create an article for us. You can do that for various reasons. You can do that because Sonnet is good for certain tasks. But the idea is you spawn a researcher and you assign a task. So the spawn agent gets a clean context from, you know, your main agent's context stays lean. So this is critical for three roles architecture. Your architect stays lightweight. Managers coordinate and workers execute in isolation. Each role gets only the context it needs, not the combined history of everything. If you don't understand anything about this, when we are actually building, for example, on our Kanban board, I talk about a concept about, you know, the builder, the the architect, and the orchestrator, which is like the manager, and finally the executor, which are the workers. So there is this three-role architecture. This will make a lot of sense in the coming sessions. But just know your sub-agents, you will be saving a ton of tokens if you let your sub-agents work in isolation.

By this time, you might have understood this is not really a beginner tutorial. We show the installation, configuring your Telegram, which is basic, done by a bunch of people. This is for people. I assume you already have an OpenTlo account, but you're only using it at a basic level. And you want to, you know, build business use cases. Grow, move, aboa. If you know for what that means, I will cover that in the skills section. I just showed you daily players. You know, here's someone who was actually running this. A guy on Unit broke down this entire monthly spend. He uses Opus for initial setup, that will cost him about $40. And then he switches to Kimiki 2.5 to Nvidia free tier for daily operations. And then a Haiku for heartbeat, less than a dollar a month. And DC for coding tests. He will use Whisper for voice transcription. His total monthly cost is about $60. And that includes Eleven Labs for voice loads and a dedicated phone number for Signal. And this is the entire breakdown that he's done. That's a fully autonomous agent around 24-7 for $60. And notice the pattern. It's exactly what we covered, you know, expensive model for setup, cheap model for daily ops, and cheapest of free models for heartbeat. And that is like eight layers in action. I actually have this post that you can read. And there are a bunch of comments as well, how other people are using it. And I would just do this entire thing, copy an entire thing and paste it in the in your OpenTlo agent and ask it to, "Hey, what do you think the best setup is for us?" And come up with your own budget. And don't get this to stop you from using OpenClaw because it's people are running this for as low as $20, $30 and getting put results. And Kimiki 2.5 is one of the best models out there. It's why it was when we link Opus 4.5 EJ. So if you know how to use that, right? Obviously, you can get a lot of value from this. And for voice, these people are spending for Whisper and Lennon Lutz. You can get Degra 200 trinites for free. And if you would just go with transcription, I think Degra, they give you like 45,000 minutes. So the crazy amount of transcription minutes for $200.

The next section that we have is security module. The thing that most people are conveniently ignoring is, I have a master's degree in cyber security. And this, it works with every and any, you know, if you are using VPS, it works, it applies. If you are working with Mac Mini, it applies. So don't worry about protocol truth yet. It will come now. It will be a thing once we start using it, doing things with it. But why this matters? There are 17,000 plus exposed instances and two unpatched zero days. And this is as of March 1st, maybe 2026. You might have changed. But Open Web UI has 31 plus known security vulnerabilities, two learning to unpatched zero-day remote code execution plus. So there are a lot of things that, um, you know, are currently wrong in terms of security with OpenFlow. Is it worth using? Yeah, if you sandbox it and you keep it completely under DPS, don't give any sensitive data. Or if you are completely keeping it in a machine like a Mac Mini or a computer, laptop, whatever you want to call it, as long as you sandbox it, doesn't have any sensitive data, it is an excellent. And what is the solution for the security issues? So two-page quick start that stops 80% of the attacks in 15 minutes. And a full Mac Mini hardening guide, which is nine steps. And complete VPS handling guide, which is 13 steps. If you want 100% security. And again, these are, I ignored some of the things that will eventually be fixed, which I know to be fixed, are not serious. We go by this Pareto principle, 80-20, to how to make, how what are the 20% of the things that I can do to get the 80% upside, 80% security. That is what I included. So if we bought CVEs, you need to know common command injection by a crafted scale. This is a big one and it's critical, unpatched. And unless you are your model is highly sophisticated, I mean, at least Opus 4.6 or maybe sometimes Codex 5.3, even yeah, other person can basically write anything. And you ask your agent to read a page and it will mistake it for a command and execute it. You don't want that. So code injection via two landing, Docker container escape, one click RC by a crafted leg. So if that went right through your head, it doesn't matter. This is not a security course. But you should know what command and why you're running that. That's the whole problem. And a lot of people don't know that and asking their OpenClaw.

Agent to do whatever and getting into all of these prompt injection kind of issues. So to protect from the obvious wants, I created something like six things that stop 80 percent of the attacks. You can read the whole thing, it's really detailed. I have prompts as well. Disclaimer: Don't run it unless you're prepared to, you know, run into some issues. It's on the surface, it looks like easy, but maybe flawed. Open law will be able to do it around 70 of these things, and if you run into some issues, it will just stop working, and you'll have to figure out why it stopped working.

So my advice: do what I'm doing right now. Everything is step by step. Six steps. I created both the site and also six things that stop 80 percent of the attacks, like Pareto principle. They have been fixing a lot of issues as of March 2026. These are some of the things that we need to do from our side. This has nothing to do with how Openglow is. It is purely to do with your setup. And also, if you want a full guide, you can do, but this is a few things that I will recommend. If we just open this VPS hardening guide, let's say, and it has about 16 pages as you can see on the side there. And every step, and if you, you know, you can be 100% satisfied and happy and content that you are, you've fixed every single issue. But for the ones who want to spend no more than 10-15 minutes and get the most out of this, this is what I would recommend. So this is more for print law security, and this is for Mac Mini. This is for VPS, this is for Mac Mini, and you can go in order. You can, you know, to get things simple, or you can, so up to you. And most of the things, this is my new Mac Mini, so I haven't done a bunch of things myself. So I can do show you a couple. And for every command, I mentioned why this is important and if you should do this or not.

Local model servers have zero authentication on 0.0.0.0. Anyone on your Wi-Fi can use your models. So this will be especially an issue if you are working from a, you know, office or, you know, what happened, any kind of shared Wi-Fi that you don't know people. And if you're using public places, public Wi-Fi. It's a huge thing. Obviously, you are not running your open floor there, so it's fine. And then who, who needs to install it? And this is the first one I missed. Only if you run Olama, Codex, LM Studio, skip for cloud only. Install Tailscale access from anywhere. Zero open ports. Who? Everyone who needs their Mac Mini remotely. So if you are using your Mac Mini remotely, which I'm planning to, so I will just copy this command and go to your terminal and just run that. And it is running. And once that is done, I want to look at what's next. So I will keep this so I can access this. What else we have? Install Tailscale, then enable firewall. Who? Everyone on macOS should on turn on your firewall. So I will copy these three commands. When C come here, and once this is done, it is at 40-42. Once this is done, I will run the next command. Oh, by the way, some of the things that we are waiting on, Openclaw agent can do this for us. I'm doing this just for a way to, get back to this. Feel free to skip ahead. So I got the 10-scale got started installed. Now open settings. Okay, Tailscale network extensions, password. And if you don't want to use remote, you don't have to. Done. And I'm running the next set of commands to turn on my firewall. So they haven't received any errors, so that's a good thing.

Three critical security variables. So you need to add a few. These are three critical security variables. Who? Everyone using Open Web UI, any model, doesn't matter. You need to run this web UI secret key. Why are we doing this? Blocks the three top three attacks: predictable tokens, open registration, and unpatched zero-day. So what, how do we do this? We need to create a generator random 32 hex. So how do you do that? You will copy this and you will run this here. And I should probably blur this or, use another key. I have another random key. I will use that, not this. And then what you need to do, this is not Docker, this is directly installed, right? So I will, you will need to do this. Export web UI key. Copy and paste. And you can paste this key, or you know, might as well blur this. And key, go. And now we need to do enable sign up false. This will prevent these open registration. And then next up, we have enable. I'm purposefully doing it one by one. So we are turning this requirement. Enable pip install formatter requirements false. And finally, we are saving everything, or we don't lose it on restarts. Yeah, I need to fix this. But and then I'm not using Docker Desktop on macOS, so it doesn't matter. But if you do, let this is you don't do this. It is, um, it let attackers escape the container and take over your entire Mac. Uh, install Lulu. And I am not doing this because it's my standalone Mac, just for, uh, Open Floor. And install Lulu. See what's phoning home. This is Mac. Well, why is it, um, what's needed? Who? This is for everyone on macOS, especially with third-party skills or tools. Warning: macOS firewall only blocks incoming traffic. Lulu watches outbound. Do not stack with Little Snitch. So Lulu alerts you whenever a new process tries to connect to the internet, catches malicious skills, compromised containers, or apps exfiltrating your data. So it is a good software to have. So a good tool to have. I will come here and again, if you are familiar with Little Snitch, you would probably won't even need this if you are already using it. But if you are not, use this. And also, never put API keys in the config. Raw API keys are easily easy to compromise. So, in this tutorial, you might see, in the interest of time, I might go the usual way because I don't want to blur and do all of these. I will just probably give it in the chat. And I started to realize, you know, as soon as I give API key off-camera to OpenFlow, it is actually calling me out. It was not a thing before, but now it is calling me out saying, "Hey, you just exposed your Anthropic key in the chat. I want you to recycle it ASAP. Consider that it's compromised. Do this now." And it's amazing, you know, how far they've come. And that's it. If you are using Mac Mini, that's what you do. If you are using a server, this is what you need to do. And full guide, 16 pages, 9 pages or something. I have separate guides for 100% security. And some of them may need updating depending on when you are watching this. But I try to give you ones that are absolutely necessary.

Provo 2. I want to hammer this home because most people skip security and click it at the later. A guy on Reddit who's been ramming OpenFlow for weeks put together his own security checklist, and it lines up with almost everything we covered. And this is what he has: Move API IDs to .env of made profit. Rotate keys every 30 days. Create .geting null for sensitive files. Input validation on email scripts. No sending without approval. Rate limit external API ports. Encrypt memory files. Use stale skin for remote access. No open RDP ports. And here's one more thing. Zeno Rochelle, the CEO of Reset, wrote a whole article or guide on giving your patient email access. And even he warns about this. Security researchers have already shown a single crafted email can trick your Open Cloud agent into leaking your entire Windows. Prompt injection is doing people and prompt injection through email content. It's not theoretical. It's not a threat. It's actually happening. That's why we covered this chapter before anything else. If you skip here, go back, and I have a bunch of resources linked. And this is real Open Cloud use cases that we have. And this is this article that I was talking about as an auto share, CEO of recent. And also we have email data leak. These are important tweets. And it's important that you have a chat with your agent. And if I do all of that, video will become unnecessarily no. It's just a matter of you going to, either you have orchestrated, that's the first thing you need to go to, or you just need to ask it why and how we can prevent this. Have we faced something like this before or be prone to facing something like this before? Cybersecurity. This is amazing. Before we moved from security, I seen big things for last. Everything I showed you, the six steps, the hardening guide, the firewall, key rotation. That's you doing it by hand. And you should know how to do it by hand. But while I'm building this course, NVIDIA just announced something called Limclaw. And I'm sure it's going to be big, but I need to show you this because it changes the game. Limclaw is NVIDIA's open-source wrapper around OpenFlow. And the entire picture's security. Every single problem I want to about. It's, for instance, a scrapped injection and container escapes, API keys and link desks. NVIDIA looked at all of that and said, we need to fix that at infrastructure level. And here's what it actually does. NemoClaw takes your OpenClaw agent and puts it inside something called OpenShare. Think of it like a sandbox, but it's not a casual sandbox. This is enterprise-grade. Like every file your agent touches, every network request it makes, every process it spawns, all of it goes through a policy day. And your agent will start with zero punishments, and you decide what it can access. And everything is long. We have this table, you know, if you are using OpenClaw, what it does. Let's say OpenClaw, your agent, same one you already built. OpenShow is a sandbox runtime policy-based security. And Neenotron is NVIDIA's open-source model. And this is optional, you don't have to use it. File access, privacy controlled, which is least privilege. It doesn't have access to anything unless you give it. Network requests, allowed lists are, you know, locked and auditable. And secrets, isolated and never exposed to Egypt. And interface, privacy router, sensitive data stays local. So remember what I said about prompt ejection through email, where someone sends agent a crafted message and it leaks your data? OpenShilver cache that. The agent can't exfiltrate data to a domain that isn't on the allowance. And it's not relying on AI to be smart about security. It's enforcing it at an infrastructure level because if you read some of the reviews or people online, they say, oh, if you are using Claude Lucas 4.6, it is pretty resistant to prompt ejection, but if you are using any model other than that, not the top-tier model, you might be just at least. And it's always something that holds people back. The agent literally cannot do what the policy doesn't permit. So even if your model is not the smartest, you are enforcing the security at the infrastructure level. And you can install it in three commands. And you can open the terminal and type and get clone HTTPS. I don't want to know this because we are doing in the middle of a session. I don't want the we can't figure everything from ground up. And it lets, I will make a dedicated video on how to set up using Neemokuro. And it will be like a 40-50 minute session. But I just want to cover this. And I believe this is a key changer for a lot of people. All the businesses who are worried about security previously will start adopting open open flow. And I was teaching or implementing this only for my closest clients and also as a business opportunity. There is this notoriety and like it's well known in this point and it has a lot of security risk. Although they are fixing it a lot, but this more from NVIDIA is a game changer. Running that comment below, install OpenShill and the NemoClaw CLI and optionally the NemoDrum models. Uh, one script compared that to six managed steps I showed you, right? NemoClaw automates all of it and adds layers you can't do by hand. And it is amazing. That's why. And it's enterprise. And to launch it, you just need to run this command: Nemoclaw space launch dash dash profile default. And now your OpenGlobe is running inside a sandbox environment. The same agent, same configuration, same skill, but not every action is policy enforced. Now, here's the thing. Limogl is still ERBY. NVIDIA is calling it an alpha release. It's Apache 2.0, which is completely free, and it's hardware agnostic. A lot of people will have this question: Do I need an NVIDIA GPU? No, you can try it on AET. It doesn't matter if it's NVIDIA GPU, AMD, Intel, on at all. Doesn't matter. Whatever you have, it works. But I want to be here with you. The manual fabric I showed you, you still need to know it. OpenClock handles the sandbox, but it doesn't replace understanding what you're securing. And while, and if you skip the first part, do at least the 80% that I was showing you if you are not into deep depth. If you skip the first part, thinking and just install no preview. Go back, you need both, okay? The way I think about it, the six steps are your foundation. NemoCraw is the enterprise near-run. If you're running this for yourself, the six steps are probably enough. If you're deploying this for a client and running it in production or building something that touches sensitive data, NemoCraw is what you're going to add and add. I have this simple table that shows personal use, six steps, takes 15 minutes. Production applying deployment, six steps plus Neenanglo. Manual happening is enough for personal use. You need automated sound web score. Production firewall rules are enough. Policy enforce status control, making it so hard just at the infrastructure level to break. And then you won't have the laws. Open short hardens everything in a production. And both of them are free anyway. And the fact that NVIDIA, the biggest GPU company on Earth, looked at Open Client and said the first thing they should drive is security. That tells you everything about why this chapter exists. Your agent is hardened whether you did it by hand with Open Client or both. You are now in the top 1% of people running Open Trust security. And that matters because everything we build from here on out, it's going to have real access to your deed, your email, your phone calls, your business. And you don't want that running on an open port with no firewall.

Now, let's introduce you to your AI. We are getting into some of the most exciting things, still a bit, explanation showing you what's possible, not into that business process yet, but foundations are important to learn first. Business brain upload. This is level one because understand this, essentially you are introducing yourself to your AI. If you don't do this step right, AI won't know who you are. It doesn't have any context of how it needs to help you, what your processes are. Anytime you have something you need to re-ask or explain, give context. Every single time. For example, imagine this: if you ask it a lead generation question, it will ask you what your product is, what your current, you know, conversion rate is, what are the channels that you are currently getting clients from? Yes, it is okay for AI to ask these questions, but imagine instead of okay, money asked me this lead gen question, I will just search for whatever is available and just give it to him. If it's being lazy, which most AI's take your command as a command, try and give you as much as possible by assuming things, you need to give a business brain. You need to upload your business brain. Does everyone need this? We have different levels. So level one is core identity, mission, values, and voice. So foundation, every agent's needs before it can represent you. So this is a level one context prompt. This is where my experience of running agency for all of those years, you know, three years of running agency. We've seen this happen in any project that we use it in. Chatbot project, voice pod project. We've seen issues after issues. But once we provide AI with additional context, it performs wonderfully. So as you can see, this is the context from template and level one. Who should use this? This guide is intended for beginners who want to avoid starting each conversation from the beginning. And what does it have? It has "About Me" section, work style, current focus, and how I want AI to help me. So what I would do is I would just copy this, right? Whichever chatbot you are using, today we have. So what would I do? I will just copy this level one template. I also give you example of how this would look. I can copy this, the whole thing. And whichever chatbot you are using now has a memory feature that it has memory accumulated over the years. And I will just paste it. Prepare this for me. Rewrite everything for my use case. I will just say, "Rewrite everything. Create level 1 context prompt for me, money." So you don't have to use Opus if that's not needed. And it will just review memory, gathering my thoughts, be right there. And we'll probably see things about personalized founder and CEO of Vertical Systems, work style is this, current focus is this. And the only reason I was able to come up with this is because my agent has a bunch of context about me. And if it doesn't have the context, you need to mention, "No, if you don't have the context, let me know." So you need about me, work style, parent focus, how, yeah, how I want AI to help me. This will become the core. Some people are suggesting that you too are brain dump. I did that, but it's not structured. What about small business owners? We have something even better for small business owners. Deep level 2 context. So level 2 is designed for professionals who want AI to understand their business context and provide more strategic assistance. What does it have? It has name, role, company, industry, experience, strengths, weaknesses, communication style, mission, products, target market, company size, business model, direct reports, key collaborations, reporting structure, team dynamics. As you can see, it's really detailed and really long too. And it's for a good reason. Now, anytime you ask OpenFlow a question, it has all the context that it needs to answer that question. So it knows what, you know, what is the meeting cadence, what is the communication tools that you are using. So it can suggest integrations, improvements, or replacements, if you will. So if it knows your strengths, when you ask for a new business venture, it knows what to focus on. And everything that you're mentioning here will serve you like a, it's your ammo, right, for whatever you're doing. And this is product, services, pricing, customer portfolios, everything that we talked about is deep context. And level three is, we are talking about large enterprise level operational blue books. It has SOPs, workflows, decision trees, escalation paths, the tactical layer that lets agents execute autonomously. So if you open level three, it has, this is the master branch that can transform how your organization works with AI 100. And this is at least 10 to 15 hours initially, and everything has guidelines and instructions to use. And if you just, if I just scroll through, executive summary and machine vision, personal profile, and it has how I want AI to assist me and your team members, team member 1, 2, 3. And also it has business architecture, product and services, revenue stream, competitive landscape, operational intelligence, marketing operations, customer success, strategic frameworks. And every meeting you ever had, all the team, have any team members, they are not restricting anything in this. We can add any and everything about the company. Current initiatives, key projects, contextual intelligence. I'm not even reading the subheadings in each. If I read that within the contextual intelligence and eat everything what's included, we might spend another half an hour or so just in context prompts. So this is why I want you to take this course seriously. You don't have to watch everything at the same time. Each thing, each chapter that you are going to implement is going to level up the way you use AI. And, this is a must that everyone is suggesting, oh, do a brain dump and it's enough. No, it's not enough. How much you can remember? Do you think a brain dump can do all of this? I don't think so. What did we do? We practice what we preach. So as you can see from here, it's around 15 pages long. And it has everything, how I operate, how what systems I build, and what my future goals are. And it has my team structure, it has my professional background, educational background, and my journey so far, and my case studies, everything that you can think of, all the products that I'm launching, that have launched. It has the whole thing. And I followed this framework to build that. So my recommendation to you is, do this to your best of your abilities. At least start with level one. Even level one will improve a lot because you are explaining what AI needs, what you want from a unit because that is important, right? You have to have a goal. If you are just reading this like a chatbot, it will be a chatbot.

And what else do we have? Your agent's brain. So these are some of the files that matter. You might be hearing about soul.md. What is soul? It is who the agent is personally, you know, values, tone, communication style. Read first on every breakup. And the single most impactful file is the soul. And then what's the identity? Identity is the external presentation, the name, emoji, avatar. I know how it introduces itself. The agent's face, if you will, is called the identity. And user.md. This is where the context forms that I was talking about will go. And who the user is, the name, the preferences, communication style. It is, know your human file. And agents.md. These are sub-agents, operating instructions, workflow rules, behavior patterns, the backbone of agent behavior, equivalent to a system prompt. So any sub-agent that we create, like earlier we created an Anthropic agent, and goes here, right? Tools.md. It's the environment nodes, local tools, you know, path conventions and aliases, risky commands, guidance, not access control. So everything lives in tools.md. And memory system. From what I've heard, OpenClaw is this amazing thing that will never forget anything. That's not true. I was looking at someone somewhere, Openclaw incident, deleting emails. So I was meta security researchers, AI agent, suddenly deleted her emails. So meta summer says she ran Openclaw on her inbox, but its size triggered compaction and lost my original instruction to get her permission before deleting. So what happened was, a couple of weeks ago, is this person, Samar, she instructed her OpenClaw agent an instruction that said, "Oh, you should not delete anything without my permission." And her command was later, she asked, "Oh, I want you to flag all the emails that I should be deleting." And that is it. And what happened was, the conversation compacts. It is not with all the millions of words, millions of context window that we have right now. Even then, AI doesn't really remember truly every single word. And it's not needed. It's not, you know, if it created code, let's say 100,000 lines of code, it doesn't have to remember package.json, right? It doesn't really make any sense. AI doesn't have to remember that. And even the chat window conversation that you are having will be compacted. So the memory system you see here, it has memory.md, which is long-term persistent memory, compressed history, durable decisions, cross-session knowledge. The agent's permanent notebook is memory.md. And the memory forward slash the date.md is daily append-only logs, loads today and yesterday's session at the start. Important stuff gets curated into memory.md over time. So you have this running memory that is happening every single day, and you have the long-term persistent memory. And if it thinks it's important in here, it will move it to the main memory, okay? So even if you consider this memory, you have to be really careful in believing that it will remember the whole thing unless you specifically save this to your memory, it won't. And it will easily compact it because it thinks, "Okay, this instruction is not important." That's what happened in summer's case. The AI completely compacted that instruction, deleted a bunch of messages, and summer was like, "Oh, Open Claw, stop, stop!" And it did not stop. It deleted the whole thing. So keep that in mind. AI doesn't remember everything, and conversations will be compacted. So this happened with me when I was using a new OpenClaw agent as well. So what's been happening, when I ask it to do something, it says it's going to do it. For example, a task comes up on Kanban board, and it will say, "Yeah, I picked it up. I started doing it." And it also creates all the sub-tasks that needs to be done, and it says it's planning. And that's it. It doesn't do anything else. So in order to fix this, I came up with an eight-phase build-out, the whole plan, but that is for development-focused tasks. But this is more of a look at a base level to really go out and build our application on either Codex or Cloud Code and completely be OpenClaw, be this brain within the system. Power, take actions, you know, read our meetings or come up with meeting intelligence or its outreach. Whatever the case may be, whatever we decide to do, we have to first make this action work. So we, instruction, it has to read and it has to take instructions from board. Why is that important? When it happens, it can take the same instruction, not necessarily from us. Next time, an agent can assign the work. I'm sure you're getting this. So, it is really important for me to fix this before I move to the build part. So I was looking at what are some of the things that I did with the other agent I was working for the first time to get to this point. Initially, when I uploaded my business brain, it was all about the user, but I didn't do a lot with the soul or identity. They evolved as we worked together. But now I'm trying to achieve the same results that I was getting with the other agent, with this agent. So we have to fix the soul and identity. So I took all the challenges that I've been facing, all the conversation that I had, trying to solve this issue, gave it to Claude and asked it to create an Open Claude agent configuration. And the issue was the soul and identity files had a lot of generic advice and vague instruction. For example, Saul was saying, "Be helpful, be resourceful." Philosophical guidance that tells the model nothing about how to execute. Without explicit operation rules, the agent defaults to acknowledging tasks instead of doing them. So this is really important that we need to fix it. Context window bloat. After 2-3 hours, your session has accumulated hundreds of thousands of tokens. OpenClock compacts old context to fit, silently dropping your execution rules. The agent literally forgets it's supposed to be building something. So we have to build this soul and identity files. And the key insight is the fix isn't better prompting, it's better architecture. Separate concerns, enforce rules, and ship artifacts. So what does the file architecture? I know we discussed this earlier, what each file has, but so who the agent is, how it executes, what it never does. And you need to keep this in a bit under 100 lines because it is loaded every turn. And what about identity? Keep under 15 lines. It is pure definition. It has agent name, emoji, expertise areas, operating model, concrete skills, languages, etc. If your model is completely new, you can add memory. If you have been using it for a while, don't worry about memory. So what is bad, what is good? I gave some examples here. You can give this to an AI agent and ask it to, I would try and not use OpenClaw itself to make this change. If you are using Claude or OpenAI ChatGPT, go to that and ask, "These are my goals." It already has a memory, so it already has a context of who you are. So make use of that data and ask it to create this file for you. Brainstorm together if it doesn't make sense because it's only 100 lines, right? 100 lines for one, maybe 15 lines for the other. So, I will show you as an example what my soul file is. And most people, including myself, didn't know how this would look like, how to access soul file for the first time. So what you would do is, if you open, your terminal, you just need to ask it to open.openclaw or open, if you know, whatever the file location is in the openclaw. As soon as you click that, it will open the workspace, openclaw first. And then you will need to find workspace. And in workspaces or just workspace, you will find identity, all the files that, you know, if you have bootstrap or agents, all the files that I discussed. And also, I think it is confusing this claw buddy integration guide with this as well. So I will probably change the name of the application to agent command or something. And let's first work on one thing at a time. If you open soul, you can see who you are. You are not a chatbot, you are becoming someone. The vibe. So I will leave this and paste this. I don't want to play with memory because memory, I've been using it for a couple of weeks now, so it should be in a good stage. So that is the only fix I'm doing, updating the soul and identity. And I will attach this guide to show you what the difference between a bad soul and identity files and a good ones. So if you run into issues, you can go ahead and change it because I am using GPT 5.4, one of the best models there is, honestly. And if my agent all of a sudden behaves weird, and that's for a good reason, that's how I fixed it. And that's how I'm going to progress now.

And then lifecycle and extensions. Bootstrap.md, first run setup interview, generates initial workspace files from conversation, auto-deleted after completion. And then we have heartbeat.md, recurring checklist that runs every 30 minutes. Keep it short to minimize token burn, only needed for background tasks. So it will check, wake up, does money have any or your user have any tasks for me? It wakes up and starts working. If it's no tasks, it will go back to sleep. But that act of actively checking is using AI, so you have to minimize. Either use model that is really cheap, really inexpensive. If not, you will have to come up with something. You have to come up with a cheaper alternative like an Olama model or something that's really cheap. Root.md, startup hook on every gateway restart, different from bootstrap, runs every time, not just the first run. And then we have skills forward slash, you know, star skill.md file and some modular capability definitions with YML formatter, installable from Plaha, are loaded at runtime when relevant. So there are thousands of skills, and this is where they go. Okay, and these are one of the examples that come to mind is constant self-improvement skill, or, you know, ability to clone skills or learn from this skill, that skill. So there are a bunch of skills that you can give your Open Claw agent. And subagent blind spot. And people, their first question when they see this Open Claw agent and the kind of demo that I showed you by not just create subagents, you have API keys, you can create one for researcher, coder, and all of that. Your agent can be main agent, but some of the cons in doing that is subagents only receive agents.md and tools.md. They don't have soul, they don't have identity, they don't have user, they don't have memory. So they are functional workers, not personality clones. And this is what most people miss. If you need a subagent to need know something, encode it in the task from our agents.md. So, most people are treating the agents or sub-agents the same way they treat agents because, they would think, "Oh, they know me, they know the user." But you see how tricky this is. We wrongfully think both are same. Agents up agents like a clone thing. No, they only have access to agents and tools.md. So keep that in mind. And also, you need to turn on agent to agent communication if you want real-time updates. What do I mean by that? So if you give a task to, let's say, your main OpenClaw agent for a sub-agent, it will send and it will get back the response, sure. But if you want to communicate, if it's a big task and you want it to communicate what's going on every step of the way, give you an update, then you have to turn on agent-to-agent communication. And we'll talk about that in the future sections. But these are the foundations that you should know and basic terminology for Open Cloy agents.

Promove 3. Memory is the thing that will break on you at 2 AM and you won't know why. Here's what one guy under it is doing to prevent that. First, you need to run this prompt. It says, "Enable memory flush before compaction and session memory search in my Open Flaw config. Set compact.memoryflush.enabled to true and set memory search.experimental.session memory to true with sources including both memory and sessions and apply the config changes." And this is really important. And it does, it changes the, it changes our configuration. Yes. And it, for those of you who are seeing, we have a bit different UI. I'm filming this on March 17, and the UI looks significantly different. They've added a bunch of cool tools like browser automation and all of that. Maybe this masterclass is too long at this point. I will make another session covering all of that. Okay, back to memory flush. What this does is it will force your agent to flush important context to memory before it compacts. Without this, Open Claw will quietly drop your most important instruction during compaction. You won't even notice until your agent starts acting like it's never met you. And this happened, the same thing happened to Summer. I think I covered this somewhere in this masterclass, how she mentioned not to delete, but it was, it got compacted, and OpenCloud started deleting her emails. Instead of preparing for deletion, it went ahead and deleted 200 emails. So what does the second thing? You know, this person actually backs up his entire .openflop folder to Super Memory every six hours via Chrome. And I have this entire table right now in this graphic. So you can see that. What does he do once a week? Once a week, he manually audits memory, all the memory files, consolidate duplicates and deletes one-time scripts. An agent has to repeat back everything it remembers so he can verify, so he knows what's missing and everything. Okay. And that's, I say obsessive or overkill, but he is a guy who hasn't lost the context in weeks, right? So the takeaway way is just back up your memory. And you can read all about it here. And 21 days ago, these posts didn't really get too much engagement they deserve, but I only included things that are absolutely important. Um, and you can go through that and, uh, do it to a level that you are comfortable with. That is the key. I need to show you something that just dropped like three days ago. And it's going to change how you think about everything we just covered. OpenCloud just released what they called Context Engine Plugin Interface. And here's why it matters. Everything I showed you, Sol.md, Identity, Memory Files, and the Backup Protocol, everything. That's the data layer, and you're writing the files. But until now, one of the biggest challenges or the most annoying things is you had zero control how OpenTRAW actually reads, compresses, or recalls that data. So the engine underneath, a black box, but not anymore. Context Engine gives you six lifecycle hooks that let you control the entire memory pipeline. So bootstrap runs when agent starts, load external data, connect drag. And we have ingest, which will fire when new context enters, filter, tag, and prioritize. So assemble, compact, afterturn. I love this last one, by the way. Prepare subagent spawn. It will set up isolated memory for subagents. But let me break down why this matters. The compaction problem we were talking about earlier, where OpenClock quietly dropping your most important instructions, with Context Engine, you write a compact hook that says, "Never drop anything tagged as critical." No more hoping the default algorithm will drop your sole file. It will always keep it intact. Or say you want your agent to search knowledge base of 500 documents before every response. For example, say you write an assemble hook that runs a RAG query against your local, you know, SQL, or if you want Superbase or PostgreSQL, doesn't matter. It will inject the top five results into the context before you even start. So your agent just became an expert on everything you've ever written. And it happens before the LLM even sees the prompt. And here's the one that matters for production. Prepare sub-agent spawn. When your main agent spawns a sub-agent, which we do a lot in a lot of use cases in this course, the sub-agent used to inherit whatever context it was given. Now you can write a hook that gives each sub-agent its own isolated memory space. This is huge. So there is clear separation. Your email agent doesn't see your financial data. Your coding agent doesn't see your client conversations or your, you know, proposal agent, any agent that you pick. So least privileged memory, the same principle as NemoClaw, but at a context level. And before versus with Context Engine, I said, no default compaction, no RAG integration, subagent share context, black box memory, hope it works. Hope is not a strategy. You will know it works. Full lifecycle control, isolated memory per subagent, pluggable RAG pipeline, and custom compaction rules. So the install is really simple. Context Engine is built into OpenClaw, version March 7th or later. And if you are on an older version, update it. If you are watching this much later, you're good. Now, are we writing our first plugin? To write your first plugin, you create a JavaScript file in your plugin directory that exports the hooks you want to use. I'm not going to code one live or use AI to build it because it's a rabbit hole. I think it deserves a standalone solution. Video on its own. But I want you to know this exists because when you hit the limits of default memory system, and you will, this is your escape hatch. So think of it this way. Everything we covered in this chapter, that's memory 101. And the files, the structure, the backup protocol. But Context Engine is like memory 201, an advanced class, if you will. You don't need it on day one, but when you're running five agents with different knowledge bases and you need like precision control, surgical control over what each one remembers, that is how you do it. So your memory architecture is solid. Let's build the dashboard so you can actually see what your agents are doing. This is where we are actually building using OpenFlow, but not just the first part. I will tell you what that means. You might have seen a bunch of people already use their OpenFlow agents to do amazing things. So this is my CloudBuddy application. By the way, if you want this, you can join my school community and fork it with one button. We have this is my place, my agents live, agent teams live, we have skills factory, and we have ability for AI asking questions, AI log what they're doing, the heartbeat, and everything, identity, their soul files, their memory files, everything. Daily memory log being saved to memory. You know, where you can open and see what they worked on every single day for me. So multiple agents, this can be OpenClaw agent, this can be CloudCode agent, or any agent that can use API, they can be here. So why am I showing you this? So this is what we call, and you might have seen a bunch of people have these Kanban boards that AI is working, and you and AI both working on things, you giving, and both AI building, or some of the things that are completely and being taken care of AI itself. So to build a board like this, or to, you can ask your agent to build this for you, and they will. But they won't be at this level, they won't be this good looking, if you will. So I found out a hack is using, if you go back to the master class, it says Flobbary Live Kit, Light Kit. So Light Kit is a light version of this. Just a prompt. And you might be wondering, why money, what is the even point of building all of this? So I want to show you like an interview that I did or a meeting that we had on February 20th. So I will just ask him, "How much revenue did this system generate for Brian?" And I will ask this question. And a $35,000 change order for a client project uncovered issues that led to 30% less jobs being quoted at expired rates. Brian mentioned he may send another $19,000 invoice for the work the AI completed. So in total, the AI system generated at least $54,000 in revenue for Brian within four days. Can you imagine AI generating $54,000 new revenue in four days? It blows my mind. And if you look at this system, one of the op-center applications is meeting intelligence. Brian's meeting, it analyzed around 30 meetings. But if you look at me, I have 1,614 meetings that AI can analyze and build things around. I can say, "Pick all the discovery calls and based on all the discovery calls, see if I missed any follow-ups or any action items. What should I have done differently?" So it has access to which day or month that I have been having meetings. Who I had more meetings with, what are the busiest days of the week, what are the least busiest days, and all of those insights, all of those things. And I can take those meetings to generate action items, to generate proposals, and that I can send clients. I can generate lead magnets directly from this. So a lot of things that you can build here. And, uh, and this is just one application. And we have email employees that can send personalized emails on your behalf. And it tracks every single thing that happened. If it, you know, how many of them opened, how many of them delivered, or how many of them bounced. Open rate, 45%. Click-through rate, 12%. Reply rate, 0%. And bounced, 126 emails bounced. And how did we get this email? And all these emails, we actually did a lead enrichment. And lead enrichment, we uploaded LinkedIn contacts just like that. And it validated at its end rate and found all of these for free, all of these leads for free. And, any application that you're seeing will add a lot of, we are talking about real business value here. So enough of the application explanation. I talked about the system, right? Now it's time to build it for yourself, your own version of this. We will build that. Now we can go back here and this prompt gives you the scaffolding, the full UI shell with all tabs, pages, and navigation. It's a starting point. We'll connect real data, integrate Superbase, and wire up agents in sections that follow. So what we are essentially doing is we have the command deck, we have agent profiles, we have Kanban board, we have AI log, agent council, meeting intelligence, premium design. So we are copying this prompt and going to any, you can go to a lovable or bolt or anything that you, any vibe coding platform that you can think of. And usually they will give you free credits, right? Five credits or so they will give you. And if you are, you know, if you someone refers you, you get 10 more bonus credits. So 15 credits, I should, we should be able to, we can probably make this work. Five credits, I'll show you. So I will paste the prompt. And if you want to change the name here, you can. I don't want to add the word light kit. And instead of Claw Buddy, you can do it in any. Right now, it is emerald green. The old one was red. So I don't mind, uh, it's being green, right? So I will just hit send. Right now it says thinking. And this is what you should be seeing. Not blank screen waiting for other instructions. So I'll build Cloverdy, a future CK, a common zone, glass morphic cards, emerald wax, and some animated transitions. Perfect. Getting ready. This is awesome. So next up we have, we don't have to wait for it. If you go back, while the scaffolding is getting ready, we can work on our integration guide. So we are building this shell, right? So what happens once you build it? You have to find a way to bring that shell into your Open Floor, to give access to your Open Floor. How do you do that?

You do that by authenticating by connecting GitHub. So you can copy this command and you can go to your terminal. And in the terminal, you will, once you run this, it says, "Where do you use Google?" GitHub.com, GitHub, uh, HTTPS authenticate with your GitHub. Yes. And log in with web browser. And it will give you a code, right? You copy this code. It is important. And you hit enter. You'll open a browser, click continue, and paste it. Continue, authorize GitHub. When you're ready, trigger a mailer. Um, trigger a mailer to verify your client entity. And I will go once to my Gmail. So, congratulations, you're all set. So, just like that, I connected my GitHub. I already logged in, but just I wanted to. So now that you just logged in with helping your CLI, your OpenFlow will have access to it. I can check, do you have access to my GitHub? And to say yes. They want me to clone anything, copy anything. Usually, if you don't want to. Now, Lovable finished generating. It looks clean. It looks, we have the tabs, we have a console, we have meetings, which is nice. It has everything that I actually asked for. What we are building. I want a visual overhaul. I want this application to be like a million dollars. We should have a sidebar menu and upgrade the header. Add some gloss morphic designs. Add some hover and glow animations throughout the application. Make a comprehensive plan for these steps. And, um, make sure this is all mob data and however many features that you can include and build in one phase. The reason I said it is because it is a trick that I use. With the last few tokens, even if you have, let's say, 0.3 tokens, you can probably build some of the amazing applications with just that one, two tokens right now. I'll see the plan and what it does. And I will, we don't even have to read the plan. It's just the idea is the plan. I explained it in natural language, but in a way AI understands, or you have to tweak it in a way that the model itself will understand. So, visual overhaul plan, million-dollar size upgrade. So I will do this, and I don't think it would take longer than five minutes and four or five credits. But I will show you what it's possible if he left with a few or at least one or two tokens. I will show you. It is done now. It has a sidebar and it has notifications. All of these are not working really. And it has a logo that's not great. But it has a good skeleton scaffolding that we can use and probably work with AI log, which looks nice for a first run and to use like a basis for what's to come. Okay, and this is a good starting point. And if you look at, we still have two left. So what can we do with these two? Now, we can take them. Be for the time being, we will come back maybe in the middle of, uh, the, in the middle of building this, we will ask OpenFlow if we should go here and build something. And then,

tutorial is not about Lovable, but about OpenFlow, right? So we will connect GitHub. And it will ask for which GitHub volume is Lovable add organization. And if you have already logged in to your GitHub in this browser, it will automatically connect it. And once it's done, continue, connect, and transfer. Anyway, it's just setting up a two-way sync. And once this is done, you will just copy this. And I have made a Claw Buddy version for us to work with. Here's the link. And I know my image is covering this. Let me just hit enter. Should we find? Usually, it's queue. Send the link to another viewer right away. And then we paste the link. And down, it will do this. By the time it finishes this, we will just ask, okay, we will build one by one, feature by feature, one by one, or something. Before we stand, I want to give or show you how to almost give your OpenFlow superpowers. How, what do I mean by that? Is being able to control browser. This is amazing. Um, I know it's not, not being taught everywhere. I know most people know this, but a lot of people don't know this too. So to let OpenFlow control your browser, we call it OpenFlow Browser Relay. So we can, so verified perks. Okay, the first thing that I would look for is, so Chrome extension browser relay. You need to first run this prompt. You need to install it. I already installed mine. So even if I do is tell the same command, V, you run this. That will install automatically. Your personal is to minus. Okay, copy to clipboard. Now it actually wherever it installed, it copied to clipboard. So if it doesn't, or it did already, we are printing that path, right? So this is the extension path, OpenFlow Browser Chrome Extension. So if you want to open it, because it's a closed folder, right? Most people don't know how to open it, or you can't even find the browser folder or Chrome extension folder because it's dot opened. So whatever path it's above there, you just paste open users, Ray, all of that, and click enter. And when you click enter, it opens the Chrome extension. This, I just need to copy the whole thing and create a new folder because it's no way to navigate to this folder. Um, so how do you set up after that? So you go to your Chrome extensions. Once you go here, Chrome extensions, you need to click on load unpack. And whichever folder you created, you click on open. And once you open, it will open up here. And when you are reloading, it will ask for a token. What do you do when it asks you for token? That those are the steps that I've written here. Also, I have a question about the gateway token. Is it something the user has to set manually, or is it something we would get from OpenFlow? So I said this, right? And, uh, it fixed all the issues. You said that manually by your environment. OpenFlow doesn't fetch one from remote service for you. So this is news for me because I thought, okay, gateway token is already, I have to query it, I have to find out, and then I have to update it. That was what I was thinking. And then I tried this, I tried that, and figured out, no, you have to search yourself. So if you just follow these instructions, you know, like, and do the D, if you just run this and set your own token, and then you run this, set your own token, the same token, by the way, and then you run all these four steps, um, you know, all of these, and then validate help. And once you do that, your Chrome extension configuration is will open up automatically. I don't want to reload it because it has a tendency of not working because you can go to this, and for me, it will say it's already configured or something. It says, okay, but maybe if I, yeah, I actually closed mine. So I'm not exactly sure if I can show you. I don't want to go through the whole process, but this is, these are the exact steps. So what happens once you follow these exact steps? They will say your Chrome extension will turn green. But for me, it did not turn green. It is still red. So, for example, I go here to email, right? I will click on this extension. I'm not sure if you can see this. I click on this extension and nothing happens. So,

so according to OpenTlaw, we are attaching it. So I want to go here. Hey, I just attached a Gmail tag and I have a GitHub email open. Can you craft a funny email about why GitHub should stop verifying me every two days? And make it funny. I know this is a no-reply email, but create a draft anyway and confirm you have access to that tab. So right now, if the steps that I did had worked, it should have access to that Gmail unless something changed. And I refreshed my gateway, reset my gateway a bunch of times in control. But right now, attached to Windows on Gmail, switch to Gmail or one with and get have open. Click on Open Relay extension. Okay, the head and click. Maybe tick once is fine. Try now. So maybe I clicked twice in the previous case. So if I open it, yeah, it is creating that funny email as we speak. Okay, what will I do? Let's see. I'm excited too. High GitHub team, respectfully, if I verify one more time this week, we're based gonna feel from Mr. Sony Chip. I appreciate the security hustle, but getting pseudo verification in a couple of days makes me feel like I'm trying to enter a secret walk in all layer in stroke Bionica. Awesome. Love that. So you can, you know, whatever model you are using, if you do that, you should be able to control your browser. And that is the browser relay. So what about ability to send emails? Not from here, not from your browser, but OpenFlow itself. And I use a tool called Agent Mail. So set up Agent Mail, get your API key, install the SDK. There are a bunch of steps. You can, it can do it for you. So you can go here and go to, and once you're in, you can get started and you can log in with your Google account. It's easier. I, this is essentially giving your agent an email. And you get like 3,000 emails a month, which is more than enough for any kind of alerts that it's doing. Oh, actually, I have an account. So I will sign in with Google. If you don't have an account, you sign up. So, welcome to Agent Mail. And you need to create, um, an inbox. So if you look at my inbox, oh, no inbox yet. So we can create one. Create inbox. We can say username, Ray, and display name, Ray. Okay, Raybot. To create inbox, they will let me robot@agentmail.to. So once I have that, I can copy the API key. And, um, can I paste? Okay, no, it's not possible. My API key. Neon API key, which I will probably revoke because it's easier. I'm going to give this. Let's do this as a test. How do I export API keys again? I want to make this full group. Okay, I was asking for a no-brainer, a simple step kind of thing. And echo export persistent service restart. We have the key, which I will, um, rebook it. So I just want to follow the good practice here. Agent Mail API key. I want to see, yeah, length 70. I did not set. Length was not 70. Okay, we'll see. We've set the Agent Mail key. Can you confirm that it works? And if it does, I want you to set up email sending capabilities and send me an email at moneymani@growthcreators.ai. It actually came back with something interesting. It says, "Agent Mail variables are not available yet." And it says, "Go export Agent Mail." And treybot@agentmail.io. And I thought it was really funny how it knew that it was treybot. The key, I did the exact same process for a co-exportation mail API key. And when I printed the length, it's 70. Look harder. And I find it really funny why it actually showed the length as 70. Yeah, it's like the harder agent when we're printed on 200. Great, then send me an email. Perfect. I received the email from Raybot and agentmail.do. So that's how you would set your browser control and the email via Agent Mail. Exciting stuff. Before we move on, I need to show you something that is very important. Everything we just did, connecting GitHub, setting up browser control, writing up email. You build that by hand. And that's important because you need to understand how it works under the hood. But there is a faster way. OpenFlow has skills ecosystem. Think of skills like plugins. There is a skill for email management, a skill for meeting notes, content creation, generating proposals, hundreds or thousands of them. And they're all open source. You can read every line, modify anything, or use them as a starting point. So you can build upon them, make it your own. And to install a skill, it's just one line. I can show you. I will open my, maybe I should go to, anti-gravity. Let me open that real quick. If I just add this, right? Npx proceed. Yes, I will show you how to read the skill before you install it. Don't do that blindly. Skills are one of the ways that you can get prompt injection. You only, you need to verify, you need to read everything, whichever you're installing. But this is to make a point, right? This is to give your agent the ability to turn any meeting transcript into action items with owners, deadlines, and you just do that with one command. And you can browse a lot of these skills from playbooks.com. If you go back to the presentation, you go to playbooks.com. And I also added something for use cases. So this is a nano banana pro skill. And it has a skill.md file. You usually read through the whole thing. And only move forward if you feel like that makes sense. Okay, you have a bunch of skills. You have and Flow skills, you should have Flow code skills. And again, the only thing that's, you know, giving stopping you from giving your agents a ton of cool superpowers, upgrades, is yourself. You see 1000 pages. And if you count this, it's, you know, there are a good 30 to 40 skills on this page. So that's 30, 40,000 skills that you can use, potentially use. And it works with Windsurf, Cursor, Klein, Claude, Code, anything you can think of, agent skills. And also, we have another site that you can go. It's called Claudeiverse, or there is ClaudeHub as well. But I kind of, you know, it has some negative vibe to it now that, you know, it was hacked and everything. Not hacked, anyone could potentially go and add a skill that is malicious. And these platforms have that risk too. And it's really important to learn how you can come up with your own skills. And you should, but take inspiration from it, read everything, and don't install anything that you don't know. And business automation use cases, productivity use cases, and email and communication automation. It has the resources and everything from X. So this is also a great resource for you to, if you're thinking what's possible with OpenFlow, what can I build? And this is a good place to start too. And, you know, this literally can give your agent superpowers in minutes. Previously, to build this automation or build a workflow, anything, make.com would take a ton of time. But now it's so easy, right? Now, throughout the rest of this course, after certain chapters, I'm going to show what I call a pro move. It's a community skill or a real-world example. That's what we've been doing. So far, showing examples of people who've been using it, uh, similarly. And moving forward, there will be sections like that too. Because if you see someone using exactly what we just built at scale, um, it does something, you know, it's worth it. People are doing, making real money with this. Think of it like an extra credit. You don't need it, but if you want to go further, it's there. Okay, let's keep building. So let's go to section six, build your own task management. By this point, every person who is doing these OpenFlow tutorials and their dog are doing and building this Kanban board. And that's for a good reason. You have to be able to see your agents, what they're working on. And you can drag and drop, and your agents should be able to move them too. It should have a proper assignees, subtasks, and this will be the backbone of every autonomous workflow. So if we look at, you know, I know there is more here, turn budget guidance. This is more to do with if you are working with multiple agents. That's why I talk about assignees, tasks, subtasks. You have to understand, you can only do so much with one OpenFlow agent. And it's just a matter of time you go start to see the gaps, the holes in it. You will need more OpenFlow agents work for you. And this is the best way, you know, what's the best way to start building your dashboard from day one, keeping in mind you are adding more, more regions down the line. So that is the key. So if you look at what this has, the board, it has five columns and one workflow. You can add more, by the way. Five made sense for me to do, doing, needs input, done, and cancelled. Every task starts in to do, moves to doing, and wait. And when an agent or human picks it up, so goes to needs input. If blocked, and if AI agent doesn't have a key, an API key that will do the job, or certain permission, a capture that only human can solve, that's where it goes to needs input. And the task is blocked. And done, which are completed, and cancelled. And also, this is a good metric to have in your main dashboard as well, because you can see how many tasks that were completed that day or previous day. So if you look at AI tasks that are completed this week, and it's only Thursday, we already have all of this, right? And, you know, Sherlock, Discord bot online, CloudBrain, and I have multiple agents here. And you can add multiple agents as well in your platform too. But we are just building. And if you missed the previous Lovable build, we built something, the scaffolding, the UI for the dashboard and Lovable. If you haven't watched that, please do, because that is what we are going to use. That application, we actually brought this to link this to GitHub. And within this GitHub, we only have placeholder data and just the UI. Our OpenFlow agent will work on it and improve on it, build these Kanban boards to the point where you can effectively work with your AI agents. So what makes a task? It has a title, description, column, priority, and due date, assignees, subtasks, and position. So this is what the task has. And the task management, you can create, you can move, you can set and update due dates, add estimate. Obviously, when there are fields, they can be manipulated, right? You can, these are the things you can do when it comes to task management and assignees. You can assign it to humans, AI agents can. Assign AI agents by agent name. So if you have a main agent, like an orchestrator and executors, these agents' names can show up there too. And you can do multiple assignees per post per job, or you can do multiple agents because each task will have subtasks. You can add more agents there too. So we have subtasks and we have AI agent integration. Agents can create tasks via API. This is really crucial. If agents are not able to use API, you don't have anything. So to set this API up, most people are using their one OpenFlow agent. Showing you how to set it up using local storage. I don't think that would work. And you need something that an API endpoint that your agents can call. And if you have four different agents, five different agents, they should be able to work together on a single platform. And that is what we are building here. And we are going to use Superbase, pre-tier. The only limitation is every seven days, you have to restart the server. But if you want to get a pro plan, it's only $25 a month. It's well worth it. And move to needs input, move to done when complete works with eight phase pipeline. So a lot of you don't know what eight phase pipeline is. It is like an autopilot, autonomous sequence. I will show you what that means in future sections. But just know your agents can assign tasks to other regions that will do every build, every task in an eight phase sequence. They, you know, then as soon as they receive a command, they don't just jump in blind. You have to set context, plan, and then, you know, create the task. And once the task is created, you build. After building, you test. If it breaks, you have to heal and then retest and rebuild. Finally, close. So those are the eight phases. And every agent that is being assigned the work, whether that's OpenFlow agent or Cloud Code agent using Agent SDK, you will have this eight phase like autopilot, autonomous task fulfillment. That your main agent can dispatch this work, other agents can pick it up and do it and give the result back. So we will build that too. So human plus agent on the same board, your board is shared. You and your agents see the same tasks. You can create a task manually, assign it to Sherlock. So if you say Sherlock or Ray, all of those things, please know those are my AIs. Uh, it is just assume it is AI agent, not Sherlock, okay? Or agent can create a task, assign you, and move it through the pipeline. Assignee system uses a unified lookup. So it is hard to remember. I got that. And also, if you want to use Lovable to get to the point where I'm at, so this is the board, right? It has all the, it looks much better than a regular Kanban board. It has assignees. It has, you know, if I want to add more people, I can add. This is, you know, you can see the glow, the hover animation that powers up, you know, when I click something, when I move something. This is spawn to agent, all of this looks really good. It is hard for Codex or Cloud Code to achieve for that reason. As a bonus, I gave you this additional UI prompt, but it's not needed. Okay, this is the most important Claw Buddy Kanban book, complete API reference. And we will ask, this is the API reference. And I want you to build this for me, right? And also, before we do that, we have to set up our Superbase. Okay, we need to go back to our OpenFlow agent. I just went through something called select a brute force way to fix things with OpenFlow. And I want to give you a quick rundown of what happened. So first, I started off with, you know, how if you watched in previous sections, I was able to clone the GitHub repository directly. I can give the message to OpenFlow, hey, clone this repo, and I cloned it and started booking. But today, it did not work. And for the longest time, all of the sessions that I was starting had that kind of access. It can run my command line. But something happened. I've been trying a lot of things this morning. And when I asked it to chrome this repo, it did not work. So I said, "Can you access my command line?" No, no direct command line execution for chat session. Then I said, "I still don't understand why is it not working? It was just working yesterday." And I closed all my sessions and opened only TUI. And even then, it's not working. So it gave me a plan. Okay, run these in your TUI. It should work. And I'm showing you 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, and 15. After 15 back and forth messages, I am now at a point, it might have worked. I didn't know, you know, I'm not 100% sure yet, but it might work. So it asked me to start a new session. If I start a new session, usually what happens was, it will probably, it will, when I start a new session, it will probably clear the chat sometimes. And I didn't want that to happen. So it doesn't happen every time, but sometimes it will. And after 15 back and forth messages, I want to see finally whether this has worked or not. Okay, uh, let's see. Run password, get status. And if it is, okay, here it is. We have a winner. We finally managed to work after 15 tries. If not, I'll have to turn some, do some more. Perfect, working ultra. So the fix was, this is what you need to set. But I will ask. So I created this document that I'm going to link in the presentation, which is what happened and how session was routing correctly yesterday and different runtime session and what we did to fix everything, kind of commands. And if you just give this to your OpenFlow agent, it should be able to figure it out. So let's begin. The entire thing was started because we had to build our handband board, right? So without further ado, let's continue. We have the repo. We will copy that repo. And the next step is actually setting up the, the next step is actually setting up our Superbase. And let's clone the repo. And, okay, it's just the UI and scaffolding. We are building every feature step by step, starting with the Superbase integration. So, take a look and tell me what needs fixing. So it's a solid scaffold, nothing broken, but it's not ready for real data yet. We are starting with what needs fixing. In North Italy, you need proper. Okay, no, wait. Superbase URL and NK. So what are the things that you would need from Superbase? If I give you personal access token, can you set everything up and use Superbase CLI? So whatever you need, just let me know what to get, and I will set, these, I will export those keys into environment. So I'm basically asking Superbase, there is something like an API key. So we need to get all of this info from Superbase. Free tier is more than enough. I'm not sure if I mentioned it, but you still need to keep this active. And if it's inactive for seven days, you won't be able to work with this. So it's only $25 for a free plan. And the pro plan and everything after that is $10 per server. So it's a really great deal if you ask me. We can do three-tier, create organization. And step-by-step, we will bring everything. So Superbase project URL is the first one. And project URL, once that is done, what do we need? We need the anon key. It's the second one. And we have project reference. And it is short. Already. And then this goes. And do not export as Superbase access token our environment, not here, database password. If you want me to remote SQL line migrations for CCLI, okay, now we need these three things. If we go back to Superbase, it asked me if we want to create a project. And I want to name it. If you want to fork my ClawBuddy project, you can join my school community. And we are even doing three-hour hackathons every single weekend. So it's going to be amazing. Every single weekend, we are going to build projects and give away that system. Whoever joins the call, or you can, it access the replay as well. So ClawBuddy is done. We have the project URL. Copy it. And you go back to OpenFlow. This is the project URL. And we need the anon key. And we have, we need project reference. Usually, project recurrence is this. They won't really even ask it. I don't know why OpenFlow is particular about that. So if we go to our project settings and keys, and we have current key, previous key, legacy JWT. If you go to API keys, we have, we need a publishable, legacy, and on and on public. We need this. Copy. You go back here. Goes here. Project ID. I don't want to add something that's without verifying, but you can see, C N F, and start ends with XDC, right? It's the same thing. Usually, they don't need it. Okay, then we also need to give Superbase access token. To give your Superbase access token, you need to follow this format, right? If you remember, we were following Superbase access token, and we are setting the token. We are loading it. We add gateway, reinstall, restart. And then we do a Python print without revealing the key. We are asking it to do, you have it. We are checking the length for it. So that is the flow we've been following to securely share our Superbase access or any API key, for that matter. And where can you find your Superbase access token? I believe it's in the account level. Access tokens. Yeah, generate new token. Sunny expires in Nehor. I'll probably blur this because I will probably use this. Okay, I can copy the token here and probably save it somewhere safe, or I will probably revoke it anyway. So I'll come here. So base access token, paste. And I will copy the whole thing. And see. And I go to terminal. Usually, it's a good practice to open a new terminal. And we'll go back here. And I will, because the session will be restarted, access token set, length is 44. So it's finally done. I implemented Superbase foundation, task board integration, and commit. And it added a Superbase client. Current blogger and machine here because Superbase command is not visible. I just verified Superbase access token in my environment. It exists. Look harder. And also try and install Superbase because you have access to this CLI. So the whole thing started because the OpenFlow said it doesn't have access to CLI. And throughout this masterclass, maybe I keep saying it for, you know, the praise. All the praise that OpenFlow has been getting, sometimes it does behave dumb. And the newer your agent, the number it behaves. So as time goes by, it will know, it is updating its memory, it's updating its soul identity and everything. It updates what it knows about you and everything. And once you get into the groove, you know, once you know the preferences and everything, things change. So I, I only set this up, this agent, around six or maybe eight days ago, max. But before that, I was using another Mac Mini. And I was using Opus 4.6. It was much smarter, much faster. And it was responsible for building the whole thing. So you can see, it's amazing. We can build workflows. Workflows, we can have agents communicate with each other. We have employees, we have skills, and we have orchestration where these agents can assign tasks to each other. And you can actually see them working on these are some of the tasks that's, you know, they deployed for each other and build when I was sleeping. That is the key. 12:40 AM, 12:39 AM. I was not even here. My computer, I was just messaging on Telegram from my bed. And it was taking care of the whole thing. So the agent will improve, and you will do a lot better together. I dug deeper and got it working. What I did, installed Subbase CLI in Homebrew, confirmed your token is available in login shell, linked this repo, ran remote migrations. So it, obviously did a lot of things. I want to see the application. Can you spin up a server on localhost? So 5137. Okay, it's blind. None of them are loading. And did we install all dependencies? Was it supposed to work? So I am not exactly sure because did not say I was installing dependencies, npm run dev, all of that is not done. Obviously, when you don't run it, it won't show. Perfect, it loaded. So it gave me issue. You got it. Dependencies installed, but it failed because environment local was missing. So it fixed it. Now we have it. So the next thing that I want to and is work on is the board. Okay, and we are making this agent in a way that is, agents can create it. How do agents create tasks? If you are only working with one agent and that is your OpenFlow agent, you don't have to worry about too much. You don't have to create the Superbase. You know, you can just use your local storage. But I am taking it, you are a kind of person who wants to, you know, have a multi-agent working for you because that is the future. And I think you are a kind of person who want to make it so you can use it with other agents as well because multi-agent is the future. So keeping that in mind, let's go to our presentation and copy our prompt. This prompt is kanbanboardapireference.md. So I will give this to our OpenFlow and paste it here. Human creates tasks for agent. So it has a lot of Superbase related things. And now you will see what perfect. This is exactly what I needed. I'll read this as a source of truth and switch our integration plan to the AI tasks endpoint contract, not direct table rights, including action update for column moves and proper to do. If you want, I'll do the implementation pass now. Current tasks data with this API and create update design. We want to make this accessible for other agents to come in and work on our ClawBuddy application as well. So, let's set it up. And mind you, you are setting the whole thing up. This doesn't exist. Are we on the same page? So Codex is taking a long time to get back. And I feel like, you know, Plot Code is sometimes even better. And actually showing what is going on. 100%. We are on the same page. You want full foundation on patchwork. I've started setting this up that way. While dashboard to ClawBuddy API contract, not mock. So base client calls. So verify proxy call. 10 stars from repo. Connected and running on this Super Busy Ally. Install project linked. Come and ship to a base bootstrap. Current state foundation is in place. Next is implementing full task lifecycle across the rest of the UI. So if we go to our task board, says loading task. Midday check and create weak memory base. I don't want you to touch or change anything in the original ClawBuddy application. Forget about that for now. Let's only work on this. And keep both of them separate. Can we update your memory based on that? So maybe I'm starting to regret the decision of making this ClawBuddy because the previous one is also called ClawBuddy. And all the tasks that you are seeing here, migrated from that app to this app. And it was so funny that I did that. And I want clear separation between them. Perfect. Let's make the app working. And the current Superbase app, clear all the to-do things, move them to done. And let's add our list of things to do in the new app. They are creating the Kanban board, testing it, improving it. And the next task is to work on AI employees, the phone agent, the email agent, and meeting intelligence agent. And after that, we go work on Alchemist, which is our SAS clone or developer. So these are the tasks that I want you to update and assign yourself for this, for everything. So I quickly want to check what we have in, what's going on in Superbase, because I want to make sure if it's real data or Ray just populated what he thought we want to see on the front end. So we have the tasks table. And we have all the tasks here. Bunch of tasks. And which is a good sign. Draft Q1 compliance report, updating onboarding. I'm not seeing all of those other tasks that were showing up there. Probably they might be something to do with the front end, which we initially suspected. He says that it's wired, but I don't see any solid evidence to that. So it finished running. And I see, hold all items through to tasks done, 60, and more, or creatively new tasks, 8. So when I go to, go here, I see around 8 records, which is true. But when I come here, I see a total of 68. I think our original suspicion was true. But I want to improve the UI first and then then work with this, because the UI, it said Lovable prompt, but it doesn't matter because it will, we are using Codex, and Codex will try its best to match that. Anyway, so even if it says Lovable prompt, I will just paste and hit enter. And it might take some time. And we'll see. So it finished the limited panel also separate from Arusha, right? The Superbase table shipped and commit. So if we open this, let's see the task board now. And we create new task, create test, this is a test. Urgent column. There is no column. And the board page, I see is completely empty. So there is a bit of back and forth you need to do. And that's common for most agents, even though they're good at coding. So done, implemented one page. And you're right, there was empty because schema pieces were missing in SuperPaste, not because of your UI. So when I go back to this, yes, it's here. I can see the assigned the names. I want you to create two tables. One is human, and the other one is agents. So for a human, you can list my name, money, and under agents, you can register your name, or make it so that an AI agent can register its own name on the platform. So we also need to start building an integration guide page within the application, where as we build that API, for example, we are building tasks. So Kanban board API documentation, how an AI agent can update the rules surrounding it, and everything. We need to do that in parallel. Let's set that up next. And you heard everything, right? What we are doing is making the application more robust and more user-friendly. That user is human or agent. Yes, setup done in parallel. And in this app, what I just implemented, new tables in Superbase, human seated with money, agent seated with Ray, supports self-registration and integration docs. Integration guide, amazing. So if we go here, we have the integration guide. We don't have money or too much. This is a test. This is a test. The changes I've been making are not being saved in the description of these tasks. Make sure that these tasks are editable by humans as well. And I'm also unable to add assignees. So the platform, the property that you're building should be accessible and it's able by both humans and AIs. So do you prefer if I authenticate myself? And it's easier to set permissions, create an account for me and give me a password as well? My email is mani@growthcreators.ai. So this is what I did when I was building my previous system because we are building systems that are not secure, that are not public, because this table can be accessed via data API, right? That's what we call a row level security. If the row level security is on, you know, only then you'd be able to, only if you are authenticated, you can see it. I still have no clue how Ray is managing to display all of this because he must have done something to make it. And this app, I just realized we must need a lot of patience in fixing these applications just by using our natural language. And I said this, you know, create this for me. And then it's a great catch. I fixed the editability issue first. Task detail panel, it fixed it. And I went to the app and I checked the Clobby application and it did some test descriptions. But it still did not fix the root issue, you know, XYZ. And click on save changes. Oh, it did actually. But I'm not authenticated. So I'm really surprised how it's working. And I went into the Superbase and I create an account for myself here. Create a new user. I set an email and a password. And I came back to OpenFlow and said, "I created an account for myself. Can you check?" And also make sure I'll be able to edit and add assignees or update them. And any new user that's created should automatically show up in humans. Yes, checked and set, um, auto-sync new users to human. And what I found on your data is my name, money, email, money@growthrate, as previously gave, looks like domain, um, type of mismatch. If you want, I can also merge and normalize to money. Yes, uh, much then. And also, the next thing that we want to fix is whenever you want to speed up the development, it is a good thing to make other regions join in because there is this is a lot of effort for one agent to work. So let's see our humans database table editor. And in table editor, now OpenFlow is about to merge both of them. And we'll see which one survives. And also, okay, merged. Kept the outlinked human display money. If I open this, is it gone? If I refresh, it's still there. Okay, legacy. This is fine. Legacy. If something else is linked there, we need a record. Okay, if I click on add, I still can't add Ray. I guess, yeah, it completely misunderstood what I mean by adding people. If I click on save, it is not saving. Assignees should be drop down of every available agent and human. And also when I make changes and click on save, as a user, I'm unable to. And also create a sign-in and sign-up page. And the setup Superbase authentication. So now we are doing two important things. One, it's fixing. But two, it's setting up Superbase authentication. If you are completely using this locally, sure, it's fine. But if you want to access this, let's say on a computer somewhere, and because we are building like a task manager, like a mission control for your AI agents, plural. Thank you. You need something that you can deploy on a Netlify site or any hierarchical version, it doesn't matter, deploy on the internet, where you can access it from a URL, from a laptop somewhere. You need to have, if anyone can log in and see, you're the only person who can access it, that is not a great setup. At least have a password to predict the whole thing. I thought we could skip this and maybe just, you know, in a future version, do a bit. But it doesn't take too long to set up authentication. Now it did two fixes for us. One is assignees are now a trap down. It goes from all humans and agents, no more free text guessing. And also sign-in, sign-up pages added. And I'm sure these pages said, doesn't look too great. We can go and check. Yeah, that's what I thought. Um, we can, we can just log in with the password that we created. And if you haven't created an account, now would be a good time to click on sign-up screen and create one right now. This is just all placeholder data. We are only working on task board. So we'll see if this changed anything. Assignees, Ray, and editable XYZ. And F E say hi because we now know AI agents can probably change this, right? But if we can edit this, probably by this point, we know AI agents can add all the tasks and everything. But I'm more into finding out whether I, okay, I'm receiving some error. Neural relation tasks violates check constraint. If my camera is blocking that, I will paste the error here. I'll say, when I try to make G, this is the error that I'm getting when I try to make new changes. Usually, it's a database issue that OpenFlow should be able to fix it. So, great job fixing it. And let's see. Right now, even before I get changes, I change this is a test. Hi, to do. I should change the whole thing. And 19, I probably need to ask to fix this. New subtasks, subtask 1. And select an assignee. I will select myself. Click on save. Perfect, it is being updated. Now that we have task board, we need to add something called task pooling. So I need you to poll this new ClawBuddy application every one minute to see if we have any tasks. So tell me this, are we using any AI when we poll for tasks? Because we need to make these just API calls and AI should only be used once the task is present. So anytime I have a new task here, you're exactly right. Polling should be pure API, zero AI, commanded flow, type 8 API call. Yes, yes, AI should only run. Yes, set this up. And once it is done, once it only does this lightweight API call, we will start assigning some tasks. See if it actually working. If see if it's actually pulling every one minute. So to make to show you that this works, I need to add a task. Something like build a to-do application, simple to-do application with so and so theme. And message me on Telegram when it's done or something. So that way we will know A. It is picking up task as soon as it comes in. And B. You can, is it an API only folder? No AI use during polling. What I added, it's already running polling every minute. Comment if you want. I'll add a smaller trigger hook. So when the count it brings work. Yes, I want you to be notified as soon as I add a task, I mean within one minute. So it might change something. I'll wait for its response and start again. By the time it's done, we can come back anyway. Test task, build a simple to-do application with cyberpunk theme. It should have visually pleasing moving elements though on hover animations. It should have priority deadline features, and use local storage for this. So priority high. Once the task is done, message me on Telegram. And also say that is done in the active chat session and give me the

local server, that you deployed this to. So this is good. Task is not like a tiny thing to do, and the deadline is today, and you need to get started on this as soon as possible. Just in case, okay, it's weird, it's not letting me create the task itself. Okay, I'm unable to create new tasks, so I might have to refresh the whole thing. So I will probably copy everything, just in case I lost. So there are some legacy task constraints requiring fields like these column key, fix, update, updated task, or create logic. Okay, great. What we will do now is we will go back to property and see if this lets us, which it did. 9 hours ago, description, assignee, I want to send me, and save changes. So if it does receive this in under one minute. So after that happened, I checked if it is correctly pulling the tasks that I'm creating, and it did not. And when I asked it to pull the tasks that you're seeing, it gave me 68 tasks, which is not true. So I asked 60 of the tasks that are not in the database to delete them. So the tasks are deleted now. So I want you to be, or give additional importance, or the utmost importance to most recently created tasks and the user created tasks. So you need to pull for that every minute. So when you see your name in the assignees, that's what you should be working on right away. When I say that, what do you find it? What is the task that you see, and what are you checking for when you're polling every one minute? Update this with this new info. I should have said, ask this question in two, or try and achieve this prompt output in two segments. The first one, what do you have? What is the current process? Okay, whatever the process is, I replace it with this new process, but I tried to do the whole thing, and once we'll see what we'll get back. Perfect, updated, and logged in. We have a total of eight tasks assigned to Ray, zero none of them currently assigned. So the assignment is not working right because I know for a fact I assigned it this task, and there should be nine tasks, not eight tasks. If you are only seeing eight, there is something wrong with that. Okay, good thing is going to table editor and go to tasks and see nine tasks. Yeah, I want to take a screenshot. Look at the tasks table and Superbase. I should be, I should be able to see what is written in Superbase. Anyway, you were right. I checked the actual Superbase task table and confirmed Polar now uses Superbase as a source of truth, same as app. It currently decides. Sign to write test task. So are you working on the test task? So next thing, I will try to do something really quick. Task. I just want you to tell me a joke on Telegram. What happens if we don't create a deadline? Yeah, it'll show up without deadline. Okay, quick task. This task, it actually moved to doing, as you can see, and once that is done, I think it will move to done. We can keep looking at this screen, but I'm more interested in and standard execution. So while I was waiting for it, I heard the news that GPT 5.4 came out yesterday, and I've been wrestling hard to find out how to use it for my own OpenClaw. And then I found that someone actually opened a pull request, and it actually merged, but the update was not out yet. So March 6th it was merged, and the available version is March 2nd and March 3rd, so it was not available. So I took a. This doesn't require a video, but this is something that you might need in future if you like something that is going on and you want to try something fast. They merged, but OpenClaw main, it's already on OpenClaw main, but it's not really on the release yet. So if you look at release, it was four days ago, and the commit was made six hours ago or so. So you can wait a few more hours, maybe days, who knows. So what I did is I pulled directly from main since it's already merged, or if it's not merged, we would have to be pulling that from to pull request the branch itself. So if it doesn't make sense to you, it's okay if you're not technical, but um, what I did was I had to update it in a different way. Not just OpenClaw update is not working. So it was still giving me the second March version, and it's not listing Open, OpenAI Codex is missing, 5.4 is missing. And then I installed it in a different way, and I installed it from the main branch, and I did all of this, and now I'm going to actually check. I actually did check, and 5.4 is available now, and I'm just setting it up. Quick start, use existing values, OpenAI, OpenAI Codex, and I will sign up with, I'll sign in with my account, continue, and I should be able to see. This 5.4 now. It's right here. This model should be a much better experience and faster experience, easier time explaining and all of that to my. Oh, this is nice. Search and moon search has a web search too. Okay, we'll get back to that. Configure skills now. Now this must be new. One, gateway is not here. So thing is, I was working on it earlier, it was not giving me an update. We will fix it. There is some additional code that we need to add, but 5.4 will be much better at this. Anyway, so onboarding complete. Dashboard, open, keep that tab to control. So I had to just reinstall gateway, and it worked perfectly. So clovering every one minute. Point down, let's hit stars and see. OpenClaw 3.3, get codex, GPT 5.4, amazing. Tokens 168 in, 35 out. Cash ahead 1%, context around 62. Too for our usage, yeah, 99% left. We should make codex work. So I'm really excited. Yeah, needs attention to do tasks are present and no answer options. So I want you to work on assigned tasks with the name Ray, and you need to do them according to the priority. Urgent tasks go first. And second priority is task creation. Oldest created tasks go first. I want to stop here for just a minute to give you a bit of context on where we are right now. So we started creating this app. And the first thing when it comes to the task board, we were trying to do a bunch of things. First, it did not have the feature to add these assignees, or it didn't have subtasks. Or even if you pick the priority, there are a few UI changes that were needed. We asked questions. OpenClaw to make these changes, and OpenClaw did make these changes. I think it did a reasonably good job. The only issue that I have with this is the time it takes to make changes is a lot. Okay, and we are only scratching the surface, and this is not. OpenClaw can be used as a wipe coding tool, but it's not its main, it's not something that I would say is at its best capability. So what do I have in mind when it comes to this building? So if you remember when we were discussing about OpenClaw configuration, especially the eight-layer to eight-layer token optimization stack, I talked about a concept called, you know, a builder and archi... So if you are following this tutorial from the beginning, especially if you watched section two, which is where I discuss the eight-layer optimization stack, I talked about a concept of builder, orchestrator, and executor. So three-layer or three-role architecture in my workflow. That is what I'm going to talk about now because that distinction has helped me a lot in, what I'm about to teach you. Because when I'm building this mission control, we finished the integration, and we are building our own Kanban board, and I want to talk about what just happened so far. So far, you have watched me struggle, prompt, and sit in the dark while OpenClaw takes my instruction, goes off to build, and comes back with something that doesn't quite work. That's not a bug in the demo. That's exactly what most people experience. And for me, I'm using Codex. And even using GPT 5.4 Codex, I've been facing these issues. Not to mention people who are using regular models that are not as powerful as Codex. They will face even more issues. When I first started using OpenClaw, it drove me crazy. I'd see people building these incredible AI systems on YouTube, but my agent kept producing mediocre results and breaking things. I thought I was doing something wrong. And I was, but not in the way I expected. So the problem isn't OpenClaw. The problem is asking it to do something it was not designed for. So OpenClaw is an orchestrator, not a builder. So it's incredible at coordinating events, running automations, browsing the web, and managing workflows. But when you ask it to construct a platform from scratch, writing complex code, building databases, wire up APIs, that's not its strength. Your instructions would need to be perfect, and even then the output is a hit or a miss. Even with the best models, relying solely on OpenClaw for construction leads to frustration. I've been there, and most people give up at this exact point. And the only reason we were even able to get to this point is because we used Lovable. And what is Lovable? I call Lovable a builder. It is a tool that's made to build. Okay, so am I saying you should use Lovable and build burnt bunch of tokens to get it done? No, you can use any of the tools, any builders that we have. So what is the breakthrough? I have separated the rules. The answer is simple. Use the right tool for the right job. Building tools. You have Cloud Code. You have Codex. You have Lovable. You have Bolt. You can use Anti-gravity as well. These are designed for construction. They understand code. They write clean, tested, production-ready software. And you use them once to build a platform. And then you never touch them again. You just need to build a platform for OpenClaw to come in and work on it. And I already showed you how it works with Lovable, just using the free plan. We just used five or seven tokens. Now I'm going to show you Cloud Code. And later, I'll even let Codex clone the repo and build a feature, just to prove that it works with any platform. So the star of the show is still OpenClaw. This is an OpenClaw masterclass. The builders are helpers, and you use them to construct the platform once, then OpenClaw runs the show forever. And that's the system, and that's why it works. So why this matters? Because most people try to do everything in one tool and burn out. They ask OpenClaw to build, orchestrate, and execute, and wonder why the results are mediocre. And this masterclass separates the responsibilities cleanly. I can show you a finished application and say, hey, OpenClaw is doing all these amazing things. Yeah, on the surface it will, but you also need to look at the infrastructure. For example, we'll be building a meeting intelligence engine, which is like an AI employee for your meetings, AI employees for phone calls, AI employees for emails. The tool itself is, you know, the powered by OpenClaw, whichever OpenClaw model you're using. But the platform, the UI, the, you know, the functionality, the ability to send all the application at the application level, the build is made by these builder tools. So once you look at the example, what we are going to do with the next few features, you will understand where I'm going with this. Trying to do everything with OpenClaw drove me crazy. You can create multiple OpenClaw agents, an agent that's specifically built for coding, you can do that, but even then you have to stay in there. You won't get the same, the same flow, the same visual updates that every step, what it's doing, you won't get that when you go with OpenClaw. Cloud Code, Codex, Lovable, these platforms are actually made for this exact purpose. So you build once with the right tool, Cloud Code, Lovable, Codex, the platform, the dashboard, the databases, and all of that, and you orchestrate forever with OpenClaw. You dispatch tasks, you run workflows, you coordinating with agents, and execute on demand with specialized agents, each one trained for a specific job. Those are reporting to the orchestrator. So once you see it in this way, everything clicks, and the frustration disappears. The system becomes predictable, and you stop wasting time asking the wrong tool to the wrong job. And what are these three roles? They are builder, orchestrator, and executor. So what is a builder? Builder is phase one, which is Codex, Cloud, Goal, Lovable, Bolt. You can pick your weapon. These are purposeful for construction, dashboards, databases, edge functions, UI components. You build the platform once with the right tool, then you never touch it again. So what about orchestrator? The OpenClaw, this is what it is made for. It dispatches tasks, monitors agents, manages workflows, runs automations, and makes decisions. It coordinates the entire operation 24/7. It is the star of the show. And most people have been asking me, you know, money, what is the difference between an automation and an agent? You're just doing automation. I see cron jobs. I see nothing different than an automation. And it's so funny for me. The way I started an automation back in 2023, if you look at my tutorials, early 2023, I was making Make and Chatbot tutorials. So that was when automation was cool. And today we work with agents. So what is if I have to give you their absolute example that makes total sense for most of you? It's clear. Automation breaks. It is like a mindless robot, or think of it like a vending machine, right? If you are hungry, you go to a vending machine, you click a button, and you get some chips or a drink or anything that you want from the vending machine. You can get it by the click of a button. So think of it like an automation. People have figured out a way when you click few buttons, you get food. That's amazing, and it got properly. But imagine. You're hungry, and in this vending machine, there is a chef sitting, and you ask him custom things. You ask for a pizza, you ask him for pasta, whatever you want, you just make the order, and this chef is really skilled, and he makes all the, you know, everything, and gives you the dish. So that is, you know, the difference between using like a building machine to get chips and getting your food prepared by a chef with this intelligent person, right? So that is the difference I see now with agents like OpenClaw core. We have these intelligent agents that can orchestrate and execute. So the phase two, which is on demand, we use executors. These are specialized AI agents that can do actual work. The researchers, emailers, analysts, voice agents, each one has a skill and the skill set to do, you know, proper job title, and a boss, the orchestrator telling them what to do. So from this point on, the way we do things is, I will be using Cloud Code, Codex for the building, and maybe Lovable occasionally to mix and match. And then we would use OpenClaw for orchestrator, as orchestrator, doing the main agent, doing things, and we will talk about the executors and AI employees towards the end. Okay, that is the flow. Okay, where do we go from here? Now that we have three roles, the builder, orchestrator, and executor, we will be building applications that are absolutely important to make this work. So for example, we are building a Kanban board because you know where your agents are working, what they're working on, and they have a place for you to give inputs, additional inputs. When they are stuck, they can move the task to needs input, and you can come in, give additional context, so they can keep working, and it works with OpenClaw, it works with Cloud Code, it works with any agent that you can think of. So what builder you should use? From this point on, any builder tool that you use will probably work. And it all comes to vibe coding, right? So when you are, we can use Cloud Code, which I'm using for my computer right now. And also we will use Codex as an example. And if you don't want to spend on these tools, you can even use a tool like Cursor, which is much better in some people's opinion. You can get a plan for $20, or you can use a Kimi code or Minivax coding if you want to stay under $20 or something. And you can even use Google Anti-gravity if you want to keep it, you know, completely free and use some Gemini models to build the whole thing. So the choice is up to you. From this point, you can use any tool to build what I am building, and everything I am building is the infrastructure where agents can live, mission control, a dashboard, a Kanban, but whatever you want to call it. This is just a beginning. You are essentially creating a digital organization and digital office for your agents to come in and work. That's where we deploy our AI employees and everything. Now, some of the audience at this point might be wondering, we were not expecting that. We don't know how to use some of these tools. You're saying Cloud Code, Codex, and, you know, Lovable and all of these platforms, we did not come here for those tools. You're right, you're not. And that's one of the reasons that I said the builder tool can be anything. Anything that you have used in the past. At the very least, if you haven't used any of those tools before, you can still use Lovable or Bolt. If you remember, that's why, or one of the reasons I styled with Lovable. You get great UIs, you get great functionality, and you just need to prompt, and it will build it for you. And I will, the prompts that you are about to see in the future in the in these pages can go or work well with any platform of your choosing. So if you want to continue in Lovable, you can. But the next question would be, hey, I already have a repository that I'm working on, and we cloned that repository and we are working on that project as we speak. What happens to that? So if you go to our workspace, and we have agent command hub, and all of the projects that it's working on. So I have this, I take this, I will just copy this and come to your main folder, and you can paste it here. And once you paste it, whichever tool that you are using, whether that's Cloud Code, Codex, Anti-gravity, I promise I will stop saying that. Whichever tool you are using right now, you can go to that tool and, for me, you can go to that tool. For me, I'm using Cloud Code, and you select a folder. And it's the same thing with Codex as well. When you open it, it will show up something like this. Or if you want to, if you go to, for example, Codex, let's say. Codex, you add a project, and you go to your main computer, and agent command hub, and agent command hub opens. And the same thing with Anti-gravity. Any tool that you choose, you're just opening that application up. That's it. Now is a good time to explain your agent better. You ask your agent, and what it thinks about the codebase. Review the project folder. I review the codebase and let me know your thoughts. I want to go to Cloud Code settings, and in Cloud Code settings, I will bypass all the permissions for now because we are doing this. Let's go back. And one of the reasons I love working with Cloud Code is I can see what it's doing. If we are using, for example, OpenClaw, I will be in this dark, you know, I don't know what's really going on, what my agent is doing. But now with Cloud Code, I can see every step of the way what's going on. So this is the overview from my builder agent. This is multi-agent orchestration dashboard, a React TypeScript front-end architecture. Task folder pulls every 60 minutes, and it's eight-minute progress reminders, enforces single work in proper estimate, how to dispatch scientists, strengths, will structure task lifecycle, and silly informalment is active, not passive, and concerns and improvements, mock data are real data split, no optimistic updates, mutations. The project has a solid foundation with good. We'll be improving this from here. I want you to deploy it into a local site server. So when we are doing this, usually what happens is it doesn't have the environment variables. I will ask it to check because the beauty of what we did earlier is we, when we used that bash commands and all of that, we were using it to essentially. So I said the dev server is live, but the Superbase is not set up. These three haven't started. They require valid Superbase credentials in .env.server. So I said I want those running too, and you can check the environment variables for Superbase, they are set. The great thing about how we use, how we gave access to OpenClaw, we exported our API keys into environment. So they are in our machine. So meaning any tool that you work later, like Cloud Code in my case, Cloud Code has access to them too. So credentials are set, and let me start all three workers. And it did, and now it's in localhost 8080. And we go here, and we can maybe sign up. I don't remember the last login exactly, or let's sign in. No, sign up. Database and saving near user. Can come here. I'm unable to sign up with a near user. So if the same thing has happened with OpenClaw, it will take a really long time to fix it, and you won't be able to see what's going on, what it's doing actually. And if you have any kind of dev background, you can quickly guide or see what the issue is. And when I was doing this with even Flawed Opus 4.6 and combined with OpenClaw, I was running into issues because I would say, go figure it out, and it will, it is not fixed. And I had to ask it, you know, what have you tried, and it would give me the list, and then it just adds a lot of additional things. Let's see, it says it's working. Let's refresh the page first. Sign up. Go to sign up. Database already saving new user, or did they already sign me in? Valid login credentials. So if we go to one page, oh, you see how, um, you know, it said, it worked, and it's not done yet. In the meantime, let me open the app in the browser to test, uh, directly. Cloud and Chrome. I don't think I gave access to Cloud and Chrome yet. So after it assured me there is no error, the error is database error saving your user. Database error saving your user. Maybe I already have the name money, if that's what's happening. So if I have to find that out on OpenClaw, it would have been so messy. Um, so I will do something. I will go to the browser and say money too. Instead, I'll sign up. Now let's work on fixing the task board first, and then we get into the next steps, building our AI employees one by one. So before we fix it, I want to change the name of the application to something else. Let's call it, let's call what we are calling the folder, right? Let's call this Agent Command Hub. Want to rename this application, change the branding, change in sign-in, sign-up pages, and the logo of the application as well. Let's call it Agent Command Hub and add a futuristic logo and update branding throughout the application. It shouldn't be called Claw Buddy anywhere. You don't have to do this necessarily. The reason I'm doing this is my agents already know a Claw Buddy application, and if I add one more Claw Buddy application, it will only add to confusion. They will confuse between the both. Now, this is the plan. Claude has prepared the app. It is currently buying a Claw Buddy with a paw emoji, and we are, I'm renaming it to Agent Command Hub. Usually, I would spend a bit more time reading it. I am not doing so this time. So it took a few minutes for completing this. It made a plan. And one of the things that I found is the environment variables, property API URL and webhook secret. Right now, my other agent will run into issues if I keep the same thing because it already knows my original already application API URL and webhook secret. So it said, I don't want to change these because it might break existing functionality. So I'm asking you to, I'm asking this Cloud Code to make, I want you to make change the environment variable names to agent command API URL and agent command bug secret, bug secret. Even if it breaks functionality, we will make the agents and we will update. He will add, we'll make agents update the integration guide with new environment variables. So I will be adding this because I don't want any confusion between the applications. For you personally, you can skip this part. But this next step is important. I want you to update the integration guide on how agents can create the tasks, update the tasks, move the tasks using API, and also the environment variables, agent command API URL and agent command webhook secret should be displayed in the settings section and in the integration section as well, so we can copy and paste and set these environment variables with our AI agents. This is important. We are asking Cloud Code to create two things. The first thing is the integration guide. This guide will serve as. If you just copy that markdown file and give it to any agent, any agent can come in, pull tasks, and it can create tasks, it can, you know, do everything. But to do that, it needs an API key. And it also needs the webhook secret and the webhook URL. So having these two things in addition to integration guide is how we are making the tools that we are building being, you know, open in. We need both the integration guide, the webhook URL, and webhook secret. All these three combined is how we are making these AI agents able to use the apps that we are building. So I haven't logged into this new account that it's created, Agent Command Hub. It looks different, and task board is, we already have the task, and it is in updating integration guide as we speak. And now we have environment variables, we have the registry, and we have living docs. I don't think living docs makes any sense yet. API endpoints, we have all of this. So we need something. Agent quick start guide, and we have the JavaScript. So this is important. If any of these are missing for your build, you need to be more thorough where you ask. I need to create API endpoints. This is missing. This is missing. So when we ran the prompt, you might be wondering, money, why, how would my agent if it's not Opus 4.6 get this? I already have a prompt for this, right? If you go to the Kanban board, I gave you two prompts. I'm not asking your agent or you to guess anything. This is the best practice. Have a task, assign the name, priority, values, and actions. AI actions also. I made the entire thing, and the loophole, and the UI prompt, whichever tool that you are using, is also here. So this is like the V1. We will have task management, we will have any subtask, and agent decoration, and that is how my agent was able to build this from what's already there. It did not invent anything new here. So if I look at here, this is the agent command API URL, right? I will copy this, or I think I can just copy the whole thing. The, what, why I want to copy this? I need to set this in the environment variables, the way we've been doing earlier, and agent command webhook secret. And if we need to also generate a webhook secret for me and add it in the integration guide, not the guide, but in the window, and the integration page. So the reason I wanted to, add these two, I think Claude has already added these two into, .env.server and local environment, but I don't want to risk it because I, it might not. And it gave me a webhook secret. I will copy it. Keep your safe because if you expose it, anyone can basically create it. Okay, webhook secret. I need to delete this. Okay, and we have this, and I will copy this code and go to my terminal and run that. So essentially, we are doing a lot of things here. We are restarting the gateway, we are setting the environment variables, we are loading them, and set length Superbase access token length 44. So I will, okay, why is it asking? Okay, so webase access token. Oh, I've completely forgot that we need to check these two, and I was checking something that's old. So if I check one of these, it should be able to, I can be sure that the other one worked. So if we go here and link 62, perfect. So I will, since my gateway started, which it should have, I'm not here, but I will say, hey, now you have agent command webhook URL and agent command webhook secret set in the as environment variables. Can you check and let me know if you have access to them? Now it will, it should be able to pick it up. I don't have to write the exact words. We have a bunch of, okay, minimal node JavaScript. Okay, I will ask Cloud Code to create me a one-click. Copy, add the, add a one-click copy button for all the integration guide. So let's go to command center or OpenClaw to see if I check harness and shell profile. Look deeper, maybe look harder. I don't know. I've been saying this, if you have missed it, if it was edited, OpenClaw misses things. You have to say look harder, or it's there. Check. The confidential profiles found, and next time I would say essential profiles. Agent command webhook not found in common shell profile situation is secret, and this OpenClaw session has not inherited current environment. Still looks on essential profile search project e4 files for URL. Verify. I'll do that. Try and find it. If not, let me know. So as long as it's not messing up with the config, I'm okay with it searching, and if it doesn't find it, I will know, and I will share it again. So I went back to integration guide and I copied the entire guide and come back to OpenClaw and when I paste it, I want to make sure, okay, it's also giving the webhook secret and the API key. Okay, it says exists, still not found. It's okay. In the interest of time, I'm just doing this. And, um, for you, you should follow the method I was showing you, okay, and don't save these to memory directly. Save this as environment variables, and also you should be polling the tasks on agent command hub every one minute, and when a task is assigned to you, work on it and let me know on Telegram. Got it. So now, uh, I'm asking you to save as unwind variables. Say they don't load up, you know, show up on chat each time. OpenClaw has a tendency to show, oh, use this webhook URL, use this secret to check. Okay, writing these two layers. And then I'm on minute poll that watches agent command hub, and then that's when a reassigned task appears. Cron, cron. So the one minute poll might be too much, and we already discussed this is not an AI task. AI is only woken up if it's, if there is something that, if there is a task, we are doing in the HTTP calls for this, then this in two layers. Perfect. Let's try whether this worked. Go to agent for mind and go to task board. I will delete this task, or maybe we have the same task. Delete does, and you does build a futuristic to-do app. Okay, and urgent column, to-do, date today, and assign Ray, create task. This because we have this one minute thing, let's see if it can pick it up. I just realized Ray was not assigned earlier. If I did not look safe, right now I assigned it properly. In go see Open domain session to see if it started any changes, but I would give it a good one minute. The thing is, it is prompted to message me if it picks up. So when it's starting build, and it's done for at least user assigned tasks for each agent, design task, it will only message me when the task is done, not when it's picking it up. So the task moved to doing, and I did not do that. So they must have, but it's not showing any either message on Telegram or any message here. So it's completely silent, but it picked it up at least, I think it did. So it's not really moving, so I'm asking. Are you not polling every one minute from agent command hub? There is a task that hasn't moved. I wasn't notified or anything. What is the reason? Checking the life move path now, whether the cron fired, cron completed, and why task didn't trigger. So the point here right now, when you're testing this for the first time, is basically see why it went wrong. Probably working with this latest technology, any tool that you take, Cloud Code, OpenClaw, or any agent for that matter, there will be breakage, there will be issues. But knowing why the issue happened will save you from not repeating the same thing. So if you look at, I'm shaking my life, but now the cron fired. What broke? One minute cron job was firing, but every run was being skipped. Status skipped after disabled. So polar existed, but the main session cron execution path never really found agent command hub. Currently returns 404 not found. Okay, the, confirm. I will give the same thing. What broke? To my Cloud Code agent. I know we are playing telephone. From this agent to that agent. But this is one of the best methods. I know for now. That's why to mitigate this, I created agent comps. So to mitigate the exact same thing, I created something called agent comps, where agents can communicate with each other and tell each other what's going on. So we can build something like that once the builder agent and orchestrator agent needs to communicate consistently. But for now, just focus on making this book because this is more of a temporary thing than a permanent thing. So while we fix the issue, I've been getting alert after I learned on Telegram. So I would say I want you to pause the polling till I tell you. So I updated a new Superbase because my Cloud Code is configured with a different account, and it has a hard time doing these changes, and I had to go to the free account and make a bunch of changes, which is not worth the time. So you can follow along exactly how you were doing, and your integration guide must have worked earlier. For me, I asked for change, and Flawed said it made the changes, and if you go to integration guide, I see, if you remember, previously was n, it starts with n, but now it's different. So I copy this and go back to OpenClaw and say, yeah, we are looking at the wrong, uh, so this project, this is the new integration guide, please update accordingly. So I added this new guide, and I want to go to my table, see if we do have tasks. Needs input. I need to make it to do, and save changes, and come back to OpenClaw and ask it to start polling, enable agent command building, and it's queued. When the message is already, when OpenClaw is already broken, it would usually be queued. So once it's done updating the endpoint everywhere, it should live unless you enable anything. Okay, had he named like that, but that's near endpoint B10. So did that, we nailed like that, but okay, let's take the same thing too. So I will copy this, and, uh, I was hoping it would be saved in the integration guide that explains it. Okay, it's function secrets. I'll have to go to a different browser because this browser don't have access to that. I did that. Try now. Okay, I said try now, and it says right now, doesn't matter, 200 OK, perfect. Why is it though? We had a task for Ray, maybe it moved to do, that's why it shouldn't say changes, or it looked for tasks that are after the change. Doesn't make any sense. Let's see. Are you checking? And the task is automatically being moved from to-do to doing? This is really unusual. So let's check what's going on. Want to make sure you have stable and it's empty. So it's a front-end issue, not the back end. So need to go to Cloud Code. I feel like the front end is still displaying the old Superbase. Let's update the task board to, and when board to link it to the new Superbase. So that is the issue. The board uses all by a shared Superbase client pointing to one project. It is now being updated to new project. That was quicker than I expected. And new project. Now let's see it. When we go back, it should be empty. It is empty. Okay, create a futuristic to-do app. And we have, hi, we have column to-do, and due date, maybe let's do tomorrow, Ray, and for some reason, I still feel like it's not up to date. I did not create my account, unless it migrated the whole thing. The authentication is not on. So I see a bunch of issues that can be wrong, and still. To check, we'll see the tasks, subtasks. Oh, and maybe display name Ray, and tasks load. And if I go to authentication, no users in the project. And yeah, that is what is weird about this. How is it getting all the humans? Okay, let's go to table editor, task, assignees, Ray's up, task board columns. Where is it? Stop sharing. How come we have the name Ray, and there is no user ID? So while we are here, I will create an account. I'll go here. We had humans and agents feature. Right now, we don't even have tables. And also, I don't think I created an account in this new database. Apparently, it cannot log out. So go to header, turn on authentication, disable email confirm sign up, and also add a log out button in the header with the profile. So bunch of issues to fix. You don't have to worry about it because we already did that in previous sessions. While this is happening, I can create an account for myself as a user. Create a new user. Again, we'll do one with what we are doing with localhost. I don't have to update it. Um, it's auto-confirmed admin and vertical systems. Task board, it should be empty. Oh yeah, it's a shared thing now. I think that if it is hard-coded with the assignees, click add, and nothing is happening, because maybe we don't have anyone. We have logout button, and it is already creating table since both clients going to same project, simplify them both client. Now, if everything is on the same project, let me just re-export main client. Now I logout button in the header. Now let me show you if I've done many years earlier. Click now for disabling in your icon. Do that. Am CP. But let me find ways with my. Okay, no worries. In your, humans, agents, tables created, auto creates a human now. So this is exactly what we did previously in the content project. Restart the dev server to bring up. It's already picked up, but I want to see what AI's that we have on. I go to table editor and go to agents. I think it is present here. Instead row name. Ray, and emoji. I just clicked save, and now you can't see with my video playing. Successfully created. I will come back here and go to task mode. Click on drop down. The assignees drop down is not working. My goal with the entire thing is, once the task assignment and works, and the agent is independently picking things off of this Kanban board, that means you can be confident that other agents can assign work to each other. You know, you don't want or you don't have to sit there and give commands if it can pick it up from the board. So that is the entire goal for this. So it says it should fix it, and Ray doesn't have a way to know that someone is actually assigning tasks to Ray now. If it works perfectly, it should pick it up, and I'm not waiting for that one minute. Is the one minute poll still working for agent command hub? So it's been a good minute. So I asked it, and there are tasks. Check again. Check the name. List all tasks. Let's see. Can you fix it so it doesn't happen again? Append it to your memory or something. So when I say append, and that is like adding more, adding this one extra line to your memory. In the past, when I was using OpenClaw, I made a mistake of asking, you know, add this to your memory, add this to your memory, and usually it does it right. But when I was using models like IQ or Gimme K2.5, they are replacing a chunks of real memory, actual work, and they're replacing with vague instructions, and it was not fun. So it says it doesn't, won't happen again. Tell me a joke. Send me a joke on Telegram. About AI agents. Agent call material. So let's keep the deadline today. Sign is Ray, and create us. So I will also change this to agent same. So we have two agent tasks. That one is more time-sensitive than other. Also, this part is a bit. I'm checking the live. It did not check until I said I have other tags, but you're not picking it up. So if you look at, tell me a joke. Opened it. Why did AI agent refuse coffee break? Because every time it tried to relax, another task came in. Black touch, and then I do circle back. Just. So this is stuck in a loop, which is bad. I need to fix it. And you are stuck in a loop. I've been getting text after, text about a joke. Once the task is done, you need to move it to done. Also, let's see why it did not create a Kanban. It's a stick to how priority work should be immediately begin starting task that is top priority. But it does done now. So if we go here, it's not moved to done. So again, that's player telephone. Go back to Cloud Code and explain this is a ship. So nine out of ten times, it is OpenClaw using the wrong API call. And the wrong API call, it is, if you are using a good model, there is a probability that the documentation itself is wrong. So I made the changes. I asked it to restart the dev server, and I disabled the confirm email. So here is an aside, just update to environment variables. Okay, it should work now. You should be able to move the task. Yeah, test moving the task and doing to needs input. Okay, if you search and go here, and you search and ask, boom, did it move another irrelevant task? Okay. And telephone. I know this is tricky, and it's taking a long time. And that is how I was able to build my original application as well. There were issues, but I was able to get to a point now my agents give work to each other and deliver on projects. So agent sends needs input underscore, but the column applied before. Scores. So I go back, go back, and try now. Try now. Try now. As you can see, a bunch of things were fixed. So root cause was tables column D with lowercase values, not column D. And then edge functions B5 is deployed with fixes. And also all five operations verified where I go. And is the integration that updated on the front end? So Cloud Code or any builder tool, unless you make sure that the code has been deployed, in the work interview, it will show character API notes. So we're running. Let's go to our agent. Refresh. Go to integration, right? And copy the entire guide and say, this is the updated guide. So now that it's switching from old quiz type, it will work. Okay, fast forward. I have to teach you something, and this is time for some lessons. I have been working with OpenClaw and Lord Code combination for more than a month now. So how we essentially use Claw is OpenClaw is the orchestrator, like I said, it only assigns work, but the work that is being assigned is for the most part, we were using Cloud Code. So I thought, okay, let's use other OpenClaw agents as the executor. So I want an OpenClaw agent to pick up the task and do it. That was my intention, and I had a lot of lessons learned when I was trying to make it work with Cloud Code. I thought it would be straightforward when it came to OpenClaw's build, but I was wrong. It took

It took me more than an hour to fix the whole thing because I was running into challenges. If you look at me, in the previous 10 minutes or so, I was fixing Superbase errors because I made two Superbase accounts instead of one. Okay, that's on me. And that was once that is fixed, I am under the impression it works perfectly, but it didn't. And I had to test it around six or eight times. The eighth time was actually the one that is working.

So we tried, you know, a Cyberpunk Pomodoro app. It failed. An anime-themed simple time-tracking app, and I said it should have 9 to 5 time slots, and you can add what you did for the day. Simple app. It's funny how it, you know, took a lot of time when I was cyberpunk to do app. The Poller is booking up and, but it is not actually working on it. It is just picking up and timing out and leaving access.

And then I decided to create a Poller and an executor, a dispatcher, or a worker, whatever you want to think. So I actually made a lessons learned document using this OpenClaw. Actually, OpenClaw made this for me. So if you look at the main problem was cron orchestration. So lessons for reliable autonomy. And what is this report and what is and is not about? Assume backend is already correct. I was focusing on wrong Superbase projects. You won't face that. The real topic is OpenClaw orchestration layer, cron behavior, main versus isolated session handling, worker dispatch execution guarantees, and task closing liability.

So the core question was, how do you make OpenClaw reliably pick up a task, start actual work, and notify human and complete the work and close the task without getting stuck in scheduler weirdness? So this is what I was facing for seven attempts. And I'm sure a ton of you out there trying to build this functionality are running into the same issues. And I'm sure this is how I managed to fix it.

The first issue, these are the three main architectural issues. And these are the workflow mistakes that I did. And here's what it will reach for. This is what we tested and what finally worked. I will try to skim through most of it because this will be linked in your resources so you can read.

Cron is basically a weak orchestration backbone. Main versus isolated session, we kept running into differences between what a main session can do, what an isolated run can do, and how much context each path actually carries. Isolated run is not equal to live main session. It does not, it won't be able to handle big tasks. The surface or runtime limits are real on web chat. Persistent worker model, we wanted to constrain by missing thread-bound session support. And session orchestration options were narrower than expected. So you can see a lot of mistakes were done. Too much logic lived in one place. Completion was not a first-class step. A stat alert was not tightly bound to execution. So it would just send an alert and, you know, stay signed and it wouldn't work.

So if you tell it to make this autonomous, here's what it will reach for. It will try cron polling, isolated agent, turns session handoffs, worker spanning, inbound notifications, and board callbacks. And all of this won't work. And this is, you should give your OpenClaw agent the URL for this and make it read and ask it not to do these things that looked reasonable but failed. We cron does everything failed. Cron detects and the main executes also failed. Architecture C, main session cron. We tried to use complete, remove Poller completely, and main session cron job. That was not also working. And Architecture D, external Poller and OpenClaw worker. This is what finally worked.

And all of that, you know, the lesson that I'm talking about, this is where the system became stable. Polling outside the LLM cron stack. So we wrote an actual script that is on the machine. It runs and it checks Superbase for any tasks. If there are tasks, it will notify the main session that is Open, which is OpenClaw. And OpenClaw worker will start executing. And also OS scheduler handles polling and OpenClaw handles execution notifications. Much easier to debug each stage. And finally, this is what, you know, is production safe. The only constraint is your machine has to be running. And it can be locked and all of that too, but you can be assured that that's when your OpenClaw is in, does life, right? So as long as OpenClaw is on, this will work, no matter what.

So OpenClaw is strongest as an execution engine, not a scheduler. Once the system stopped asking OpenClaw cron to be a durable, always-on orchestrator, and instead used OpenClaw for what it is best at, which is agent execution, messaging, file work, and task completion, the autonomous loop finally became reliable. So the real lesson is not OpenClaw failed. The real lesson is reliable autonomy comes from giving each layer the job it is actually good at. So that is what we figured out.

And for the final one, remember how I said, like a time tracker, 9 to 5? This is what it created. And it took less than, it was really fast too. You can, and the final one, I asked it to create a Pokemon-themed or something Pokemon-themed 2D pixel to-do. And if you look at the Telegram, it actually implemented the last one. So this is the to-do. 11:27 AM it picked up, and by 11:29 it finished and sent me a message that is done. So you can look at what it created.

And yeah, this is how an agent Kanban board works. The idea of this, the main goal for this is other agents can assign agents work, and agents will pick it up, work on it, and notify us. So this is the key that has been missing for a lot of people. They are essentially clueless how AI agents can work on these tasks. Yes, you can ask every single day, sitting here and typing on your Telegram, but you need to be in a place where you have an intelligent agent that can handle and make other agents work for you.

This is one of the most exciting topics that I am interested in, and why I made this master's in the first place. And also realized why there are not many OpenClaw masterclasses out there. Maybe this is the only one. And this late, because people, we have two people, both the extremes. And people who are giving away 30, 50, 100 use cases for OpenClaw, and they will just give you a bunch of text and say, "Go figure it out." But actually implementing that is the hardest part. And there are a bunch of other types of videos who do cover the setup. They go into, "Hey, this is the problem. You need to run this. This is what you need to do." There are setups like that, but there are only, you know, most setups cover one stage, maybe two.

But what I'm covering here is the business operations tool, and it has seven stages and zero gaps. I have been running my AI agency for three years now, and I've been documenting my entire journey here on YouTube. And every feature that we are about to build in this masterclass moving forward maps to a real business stage. We will do outreach-related employees, will do discovery-related, proposal-related, onboarding, delivery, intel, retention. And this isn't a tech demo. It is an AI-powered business engine.

So if you look at the table that I prepared here, it has outreach. We talk about cold calls, email campaigns, prospecting at scale, AI reaching out before you wake up. We will build all of this. Am I saying there will be outreach agents like three agents and four agents? No, there will be overlap. So if you look at Lexa, the voice agent, you can use Lexa in other places as well. Or if you use Nova, this email agent, we can reuse that email agent here as well in retention. So usually, what I'm talking about, the features that we are building will have some overhang. But we are creating every step of the business.

So, discovery, ability to research prospects, prep for meetings. Know everything about who you are talking to. All of that meeting intel, research hub, and cognitive memory is covered in the second area. And the third area of proposals, you can generate quotes. You can generate statements of work, deliverables. You will be able to turn the discovery insights into revenue documents. And on boarding, we will talk about how to welcome clients. We set up the workflows. We create dashboards. Because first impression is best impression. Fully automated.

And delivery, build and ship actual work. How does that look like? After you have the onboarding call, we will see a discovery call example where AI will identify that the discovery call that you had is something related to sales, and it will that and makes you automatically, without asking, a proposal and a meeting, your prototype that you can show to the next meeting. And it can also, if you let it, follow up with the person who is. I actually booked a meeting with you for a follow-up point. But we say that for the call, it's an advanced feature. But you get the idea. We will apply the same principle in the TED degree also. Having the onboarding call and making the AI agent build the prototype, the team can then come in and use for the actual work. We will see that.

And when it comes off, when that is done, we have agents dispatch, building, validating, and reporting autonomously. And competent intel, you can monitor your competitors, you can find content opportunities, work opportunities. You can stay ahead of your market. And finally, retention. You can do follow-ups, you can do check-ins, upsells, renewals. You keep clients happy without the manual effort.

So, this is the frame. Every section that follows AI employees, multi-agent, command center, automations, maps directly to one or more of these stages. You're not building cold AI tools. You're building an autonomous business operation system. Every feature will have a job. The most favorite part of my entire masterclass, build your AI employees. These are full-time autonomous workers. And they, they prep full goals, even campaigns. I YouTube it down each warm and dedicated employee. None of this take days off. It gets its own Alpsender page with data. And I need the example close. So one employee handles meetings, another runs outreach, another watches your competitors, like 20 words. I think no day. So I am really excited for this.

First, we will start with meeting intelligence. I am using Firefly, but you can use any application that you have with it. And that's Firefly's order as well. As they have an API key, it will work. And then Lexa is a full agent voice AI for both inbound and outbound calls, appointments, follow-ups. And Nova is an AI email agent with campaigns, templates, analytics, operator price. And create a command to. You'll have YouTube Banneria text, competitor intel, content calendar, and outline detection.

So for all of this, first, before we start an employee, I will show how it works in my actual property application. And then we will build in our agent command center. And then we will build in agent command hub. So this is the meeting intelligence that I have, one of the ops and application you will look at. Meeting intelligence, this is good action. It has all the meeting data, all the meetings that I had. It has, as you can see, a lot of detail, how many hours, what is the average meeting duration, action items. And for all the day I signed up June 2024 to today, how many days are in having every single month, and which days are my busiest days of the week, what are the external organizations that I have been having needed. And every day they will have detailed takeaways, topics, and everything to go to an actual page where it has the big data. I don't want to, but it's there. I want it.

And after from each leader, they can take automatic actions, or we can do actions that are not automatic. So if I say this is a proposal, let's say if I go to service calls, discovery calls will show up here. I can come here and you can't proposal, it will do the proposal thing. In fact, lead magnet. And if you want to automate it, which here is, if we go to automations and in automations we have functions and webhooks, we can create a function which we say crosses the following data, and we can detail HTML. But if you look at the webhook, we created an endpoint. And in our, pardon me, if you will get the clorned. I have ETID and webhook secret and webhook here. And any time the meeting is done, it gives me transcripts of it in action items, and they are sent here. And I can see that in webhook, and this webhook is processed and comes here. And once that is done, they will be in processing. And less than a discovery call, team meeting, a steam meeting, it can email. We already set up email right, using agent mail or something. So if I want to email my team members after a meeting is done, I would say, "Hey job," and these are the action items that I want you to book on my agent. And then over. And the same thing with one-on-one calls. After I have a meeting with the client, that trance that can be operated again. This is not generating AI stop. We are using our own knowledge piece. These models are prompting in such a way they have all the context of not, we know the best factors in everything. And they generate the report. They generate the proposal. So, every proposal improves the pro because of the feedback that we only proposals that are going converted into actual sales. We learn from them and we try and avoid things that we did for field proposal as well. We learn what we did differently, to what I done heard this script was. So, we have everything.

So this automation feature is amazing. First, we will build our meeting intelligence, and then we will make an Open provision. As soon as the meeting comes in, if it discovers it's a discovery call, it will automatically pick it up and create a proposal and create a prototype. That is the flow that we are building for VD intel. But the full version actually has a lot more. Then you can get this in an agents in the box community, a school community where I will be doing three hours trackathon every single week. We've been built together. So I would love to see you later.

Now, to build our meeting intelligence assistant, we click on this page, and it has, you know, the three data types and one dashboard, meeting records, overview KPIs, and calendar events. We'll be building all of this using our builder, which is CloudCode. You can use Codex as well, or anything, any builder tools for that matter. So what does this have? What is the workflow? Workflow is we use a Fathom webhook. You can use whichever webhook, whichever email, you can use whichever meeting assistant that you use, Fireflies, Otter, doesn't matter. It will have a webhook option where we can create a webhook endpoint using our Superbase. And once we create that, it will start receiving any kind of information from any app that you want, like you've seen earlier. And raw reports, these will take in raw reports, which are transcripts and action items, and any summary that you, that whatever your meeting assistant platform is sending. And we will transform, we will extract these fields, and we will create ops data. We will create meeting records and save it to our database. And it will be a searchable UI. You will be able to see all of that in the dashboard. So these are the fields that we are taking right now. For any reason, your Fireflies doesn't have any of these options, you are welcome to change. And your builder, whichever agent that is, once you start working with the webhook, it will tell you, "Okay, for example, meeting type is not listed. Can you guess what the meeting is based on what it says?" If it is between two people based on title, it will mark it as one-on-one. If we say it's a team call or Q&A, it will tag accordingly. And if you see anything related to discovery call, sales call, anything related to sales, it will tag it as a certain thing. So can you, these are the things, these are the type of meetings that Fathom or Otter may not produce, or Fireflies will not give you out of the box. This is something we are classifying. There are fields like that too. So it is completely up to the builder and yourself to how many fields that you want to change this to. And what are the UI components, KPIs and charts, search and filters, meeting cards, and feature pipeline? Prerequisites, you have a meeting recording tool that sends webhooks. Supported options, meeting recording Fathom AI, Firefly Sorter, or any tool that exports transcripts and action items via webhook or API. So calendar sync is optional, Google Calendar via Make.com. So I like to have this feature where I don't like to give my calendar access directly to OpenClaw. I don't like the idea of connecting where OpenClaw can create events. I create something called a buffer in the middle where I only give my Google Calendar access to CloudCode. So CloudCode is this middleman. Anytime a change has to be made, OpenClaw will notify my CloudCode agent. My CloudCode agent, if it fits what I tell it to within the set criteria, it will create an event for me. So I don't like giving grains or control to OpenClaw yet. So that's why I use in this multi-agent orchestration, multi-agent setup that I have, CloudCode has control. And what else? Building meeting intelligence from scratch. This is the prompt we will have with our builder. You can use and try and give this to our OpenClaw agent as well. So if you see here, we are talking about build and meeting intelligence module inside Claw Buddy's Op Center. This is a classic builder prompt, but you can try and wrestle with OpenClaw if you want to, but it will drive you crazy, trust me. You will manage to make it work in the end, but it takes a lot of time and you have to do a lot of additional work to make it work. A simple change that right now it is building this application. If you see here, I said Claw Buddy's Op Center. I forgot to change the name, FYI. We are making changes to Agent Command Hub, not Claw Buddy. And also, let's create a page with the title Op Center, and the meetings will be one of the cards in this Op Center page. And when someone clicks that page, they will see the existing meetings page. And Fathom meeting key, Fathom API key is not available yet. I will provide it real soon. Let's build this meeting intelligence first. So this is what I said. And my job now is to go to the API key Fathom. I want to create an API key. Agent Command Hub. Create API key. I will reboot this key as soon as we are done, and I will make a note of say, I saved the API key. Right now, I'm unable to see again. I can see this webhook secret everywhere. I can copy it, but I cannot copy the API key. So where do you paste the API key? You might ask. So you have to go to the anti-gravity or any IDE for that matter. You can use, if you use Visual Studio Code or Cursor, doesn't matter. And there is something called .env.local in Agent Command Hub. You go there and you paste your API key. This is the place where you will have all of your API key. This is the environment variables on CloudCode. So I will save it. I will come back and go to CloudCode and say, Fathom API key is now set as an environment variable. And I'll just leave it there. It is already working. So good. Now I have the full picture. Let me build navigation content. So Op Center cards can navigate to sub pages. OpCenter.dsx hub page with meetings card as placeholder for future modules. Okay. So I think it completely misunderstood and it will create an Op Center. Okay, meetings is now gone. So you can see meeting intelligence as calendar sync, action taker, proposals, and follow-ups, analytics, and reports. Okay, we have it says that we have ten meetings, and these are all placeholders. Now that I've given the API key for my Fathom, it will try and get. I want to make sure we only ask to bring in this month's meetings or this quarter meetings. I don't want you to completely fetch all of the meetings. Only bring last hundred meetings, no more than that. So now it's using the API key and fetching the Fathom 10 meetings from December 2025 onward. I have full Adem Fathom API shape. That's not true. Maybe I'll have close to 100. 10 I had this week. I copied my Fathom API key and I'm going to Superbase and paste it in the secrets. FYI, Superbase now has Fathom API key. Also, I like naming my AI employees. So, we have a name for the phone agent, which is Lexa, an email agent Nova, and we can name this meeting assistant as well. We can call this, let's change the Op Center card name for meeting intelligence. Let's call it Jane Meeting Assistant. So I did something crazy. The Fathom API key secret was stored in Superbase with surrounding double quotes, and that's why I was giving a 503 error. And CloudCode was able to figure it out in like a couple of minutes. If I was doing it, it would have taken me a long time. OpenClaw would try and fix and running in circles. That's why I believe a builder agent is so much better when it comes to building these one-time tasks. Okay, next done. Updated the card title in Op Center and the label in the headers now says Jane Meeting Assistant. Right now, the task board is working. Op Center is working. Now we have our Jane Meeting Assistant, and when I open it, it says loading meeting from Fathom, and it has my meetings that I had with Brian, Q&A, Oscar. So it brought in 10 meetings from past 3 days. That is okay. I asked for 100, I believe. Um, but if we want, it can get us more. Okay, has monthly trend. It has external. Uh, it looks good. Has action items. We'll show who wants with action items. It doesn't have actual items. It it removes them. Okay, this is good. This is exactly what I wanted. So the next question people obviously ask is, what is the value here? Money. What did we just build? So now we have to set up one, the automation. As soon as a meeting finishes, it has to come here. That's one, the first thing that we want to do. And the second thing that you want to do is build a function. What is function? A function is a place where you set what happens with this raw webhook data. Do you want to take this webhook data and, you know, ask it to ask an AI agent to review? That's one. And it's a discovery call, assign to an OpenClaw agent, and then ask the OpenClaw agent to create a proposal and ask an OpenClaw agent to create a proposal, and also a prototype if you want. So whatever you want to do, you can set it up. That is the beauty of this. So I will come to my CloudCode agent and talk to that. Right now, we have two options. One is to create these webhooks feature within the Op Center or within the meeting assistant. We can create a new tab and create webhooks and functions. So we set up a webhook endpoint, save it to Fathom, and as soon as there is a meeting, you can, or you can, you know, if you are not using one of these features, let's say console, I don't want to use console. I will say I don't want to use the console page. Let's rename it and change it to a page called Automations. And in this Automations, they need two tabs. One tab is called Webhooks and Functions. The other tab is called Webhook Q. Webhooks and Functions are interlinked. You can create a new webhook. And once the webhook is created, you can create a function that goes with the webhook. In this function, you should be able to specify a prompt and the raw data that is gained from the webhook can be processed using this function and the prompt. For example, if we build a webhook for Fathom, we can add that webhook to Fathom, and Fathom starts sending meeting data to that webhook. And in my function, I will set in my natural language, "Explain, process this data in so-and-so format and send it to the meeting assistant page." And you need to also update the table, the meeting table accordingly to accommodate this new process data. Okay, I should have read what I said instead of. It's not console, it's council page. So OpenClaw, when you're working with it, usually goes to queue. But this CloudCode, it's better, and it actually checks before going through the whole thing. Nine out of ten times. So I got this question for webhook queue tab, what should it show? A log of oncoming webhook payloads with their processing status. Yes. And when a function processes webhook data via air prompt, where should the processed data be stored? So create another tab saying Reports or Results. That is where the processed data from webhooks layer. And also I can make specific requests such as this one where you route it to specific tables. For now, the Fathom webhook will only go to the meeting table. But for others, it might be different. But create a generic table and a reports page. So should the AI prompt processing happen via LLM API, Cloud, or OpenAI inside the edge function, or should it use simple template-based processing for now? So make it optional, fallback through. Okay, API keys, uh, consent. So I would like to use an API call because it's easier. Um, because we are always going to receive templates based. For now, use structured templates and rules drawn. Something can add LLM later. So yes, let's use that. So for my other meeting assistant, what we did was, we get this, and it do not have any kind of processing. It will just stay there. I can go there and add action items or proposal or lead magnet. Whatever option that I pick, it will then start the process. It will create a task for AI on the board. Then OpenClaw agent will come in and do that work, or CloudCode agent will come in and do that work. So that is how it was set up. And that's why I said no API key, just this. I want to build that. I want to actually make this much better than my version by adding automated processing. That's why I'm contemplating whether I should add any API keys. So it says user wants to remove console page, replace with Automations page, and we have a bunch of text there. And now it's compacting conversation. It will be a while. So our page is now ready, I think. So let's see. And also, when you are working with Fathom, it is a bit different than other webhooks. It will give you a webhook secret, and your builder agent has to deal with that in a different way. I will show you what I mean in a minute. But for now, I will go here, go to Automations, and no webhooks yet. Webhook Q, and Webhooks and Functions. Create Webhook, Fathom Meeting Webhook. So let's create it. I'm unable to create new webhooks. And when I close it, it's not creating. But I actually create, click on create webhook, it's not working. And also for Fathom only, we need to create a webhook in a certain way. They have a webhook secret. Can you go through the API documentation and let me know what's the right way to fetch meetings? So, as soon as the meeting is done, we will receive that webhook. So when we go here, manage, and click on add webhook, it says destination URL and it has a secret, right? So what I want to do, it has a webhook secret. That's what we told them, told CloudCode. So let's see. Okay. And now, when that is going on, I wanted to add something else. So, if you look at here, we have something interesting. You can see a bunch of options here, which has soul, identity, user, memory, agents, tools, heartbeat, and daily memory logs, right? So your agents have all of this, and you want to keep track of these things, and it's important. I'll tell you why, because if you find things, if you keep a lot of things every single day, especially if you are using CLI, it's so hard to read what's going on. You cannot open memory.md or you cannot open daily memory logs just to get information to find out what happened that day. That's really unrealistic. So what I would recommend instead is creating a page on whichever dashboard you're using. So for example, if you look at here, we are not using multi-agents anyway, but we will towards the end. We are using task board, AI log. We will make AI log work whenever our agent does something, whenever it picks up a task, whenever it is doing heartbeat, it has to update it in the log. So we will work with that. And the automation, right now we are working on it. Any employee that we are creating will go to this page. Integration guide is also functional. And as soon as we add more things here, we will update our integration guide. So our OpenClaw agents, CloudCode agents, any agent that you create can come into this application and use it fully. Okay, that is the goal. That is what we are trying to do here. So I don't think it will work because, uh, maybe it is opening up in my actual computer, the one that I have on this side. It won't work. So it will fail anyway. Port 8080 is what it says, but let's see when, if it has access to what's going on. Have an easier time. So we have Automations, and we create webhook, Fathom Meeting Webhook. It's okay. Now I will delete this. Create new role level. While it's, for table webhooks, I will copy this. Gives me an error when I try to create new webhook. So usually when CloudCode works fully, and which most of the time when I'm using it on my computer at least, is it has complete control on browser as well. So it goes, it goes to the page, and it looks at all the browser. It will click through the web page, scrolls down, creates web hub, this everything. So that is a much cleaner flow. You don't have to test it yourself. Now it's just kissing and policy exists. So it actually finished working. And if you go to our 8080, I'm not sure if this work without refresh, but we'll try. Okay, Fathom webhook one function. So the meetings are the meetings. Your webhooks. And so I have a webhook secret. Where do I add that? So my Fathom meetings are automatically fetched. Where do I add the webhook secret? The API key is in the environment variable. Check that. I want to set up another automation where whenever you receive a meeting that says discovery call, I want you to send that meeting, automatically create a task with the description asking it to create two artifacts. One is a visually pleasing HTML document with a proposal, with its own estimated placeholder values for the contract value. And the second artifact is creating a prototype, a functional prototype, with whatever the client asked. So it has to take a transcript and create these two artifacts. I want you to create a function that automatically does that. It will add this prompt and adds the entire transcript in the description and assigns Ray and the deadline is the next day. If today is March 11th, deadline would be March 12th, and the priority is urgent, and it is, it has to be in to do, and the assignee name is Ray, R-E-I. Let's set that function up. So right now, it already set it up as whenever I have a meeting from Fathom, it will automatically send it to this webhook upload gold. Actually configured the whole thing without us doing anything. That's amazing. And the second step, what we are doing right now is we are setting another function. Update. I need to do three things. Update the Fathom webhook to include transcripts, currently disabled. Add new processing function for discovery calls, task creation. Update the edge function to hand them creating tasks with transcript. So three to-do. And once that's done, what I would do is obviously we cannot simulate a discovery call now. So I would use one of the past discovery calls. Let's say we go back to my. So let's hit refresh. Yeah, it already configured on its own. My recordings, transcript, summary, action items, perfect. So we'll go here and go to. Discovery. So we go here and we can pick a call. Maybe share link is enough. Um, what I will do is go here. Once you are done with setting that up, I want you to simulate this function. Assume that you already received a discovery call, which I'm sharing the Fathom share link. Use your API, Fathom API to fetch the transcript, action items, and everything for this call, and create a task for Ray saying that he needs to create both HTML proposal and the prototype. So discovery call to task automation. As discovery call payload, if checks meeting contains discovery call, case insensitive format transcript. So it created some functions to figure that out. We are not using AI here. And then we are asking this. So done. I'm setting that up. I want you to simulate this function. Assume that you're so it is fetching their transcript for that. So we'll see now. We'll go back to the task board and. We will have Ray here. I don't know if this is going to be dynamic or anything, but we'll see. So first, we have this, and she has written, yeah, I thought it was supposed to and to pick the meeting name. So it, uh, I should have given the transcript itself, but it's. Bad escape character. What it's doing is it tried to connect to the browser on my computer, and even when I gave permission to connect, it didn't. I'm not sure why. So now what it's trying to do is, since the user said to simulate, I'll draft a realistic Fathom discovery call payload and post it directly to webhook ingest endpoint. So I'm completely okay with that. While it's doing that, I will come back here to see discovery call follow-up task. So it created something for us for Ray. If Ray is picking it up in five hours, urgent. So we'll see if Ray picks it up. If he picks it up, it has to first send me an alert, and then he works on the task and gets back to me once that's done. That is the flow. Why are you not picking up the task? So, I feel like I know what's going on. There is a thing called user assigned task and agent assigned task. When I was. Okay, codex error. Okay, there is an error apparently. Okay, we'll see. Okay, you know the task. So, when I was using multi-agent Discord server, what happened was when a user was assigning work to another agent, it did not pick it up. So when I asked for a reason, it said, "I did not pick it up because it was agent created. We only respond to human created tasks." So I thought it was a thing. Carries an item because there is no actionable rate as to know they run has no action posts. So I will go to my CloudCode agent and ask it what the hell is going on because this is the crucial piece. If you know, why is it null? If other tasks has my user ID, it should have my user ID for this one too. Okay, it should match the other tasks that are the tool web and everything. If it has my user ID, this newly created task assigned by you should also have my user ID. Updated and let me know. So this is exactly what I'm talking about. If it's not assigned by the user, they have a hard time. Okay. I'm really surprised to see this name, which is my first, really first agent that I built when I was doing my OpenClaw video. So that was in January. The entire thing that messed the whole thing up. Must be this, right? So let's go back to our agent. Okay, it's gone to doing. So let's, if you open Telegram, discovery call follow-up, discovery call, and make up website redesign, doing urgent execution started. So it might take a while because we asked two things, one is proposal and one is prototype. It so I will. Okay, HTML over it and, you know, it assumes. So, okay, sorry, I was reading after the task is done, it will give us something like this. So I will wait now. The time is 7 p.m. Let's see how long it takes for it to do the whole thing. If it does, and before I pause, I want to check whether it made any, if there is any change in here. Now, as you can see, the agent here is completely silent. It's not moving. We are spawning another broker completely to do that task, and it's working in the background. And there is a good chance, based on how big the task is, it might fail. So the key is balance. And also, how long have you set up this timeout for? So if it fails, we will inquire or we will see what the issue is and we will probably fix it. If it works, it's great. By the time it's done, we can probably work on other things. For example, we need to create a new page. It is called identity. I'm adding an inspiration from another application. It should have all these fields. You should also create a daily memory log table where agents can create memories every single day, and it should be agent specific. Right now, Ray is the only agent. You can be an agent too. You can give yourself a username and create an identity on the application. Let's do two things. Let's create identity folder and cards for each available agent. The first agent is Ray, so I need to create soul, identity, heartbeat, and all of that. So Ray can populate his entries. Ray can populate his daily memory log and everything. So let's build that next. Once you do that, I need you to update the integration guide. So what are the API calls to update soul identity, heartbeat, upload daily memory logs, and everything? So we are doing two things, right? Let's go to my desktop. Add a screenshot. And this is it. And we are at the corner. Let's go to our Telegram. I received an alert. Completed discovery call follow-up code, etc. An error record processing your request. You can try your request or contact us. Message sequence number. I'm turning the transcript into two artifacts, and it again polish HTML clickable. Then I'll smoke test commit mark has done. Both artifacts are created. I'm smoke testing the proposal type locally. So I want to go to this page and actually, let me open OpenClaw and see what is going on. I don't have OpenClaw. Okay, let's go to Domino and open this and go to workspace and Acme redesign proposal, proposal.html, website redesign, 3D product showcase launch target. It has executive summary, has phase one, core marketing website, 3D product showcase, client portal, Salesforce CRM. Okay, phase three. If I asked PDF, I'm sure it would have created a good PDF. To Codex's nice. Let's see. Prototype, marketing site, and line 40, live demo data, marketing site. I don't think this is really that good a demo. So you can see we can let this create. Client portal, refresh them all, Salesforce linked account, open dashboard. It's not really working. So if I say we need a better way to give feedback and set this up, but this is how you can automatically assign tasks to agents as they come in, whether that's processing, creating these proposals, or sending messages to your team members, whatever the task is, they can do this for you. So this is a good experiment. And the way it was prompted, creating these two artifacts, there we should have been more like, okay, after they are done, you need to upload those both HTML files to the automations and automations folder reports. So I will actually, once the identity is set, and let's refresh if it's done. Do we have any questions? Did it ask us any questions? Okay. It is still building. So I just got an idea. So we can ask. Once the task is moved to done, the created task, we need to add another task saying the created HTML reports or files should be uploaded to the reports section with the task name as the title. So create both the table and the API call to upload to that table HTML file within the task and assign Ray. So when you have a new agent, let's say you wanted to create a CloudCode agent. So this is what we are going to do. We are building this meeting assistant like a raw meeting assistant using just CloudCode as a builder and OpenClaw as orchestrator and executor, right? But in the next tutorial, in our CloudCode masterclass, what I'm going to teach you is that we are going to do this completely with CloudCode agents. Maybe we will use OpenClaw, but CloudCode is so much better in creating these prototypes and everything. So towards the end of this tutorial, I will also show something called the eight-phase autopilot autonomous builds where OpenClaw agent will assign work to a CloudCode agent. And that is the final chapter that I have planned. But for now, just understand the principle. Okay, just understand the principle of how this works, not really the final output. It will be, there is a lot of room for improvement based on how we are prompting. For example, my proposal, I did not give this agent anything about what a good proposal should look like. But we have three-page long prompt for our proposals, the way we write proposals. So once I give that prompt and create a skill around that prompt, everything changes. For example, if you look at my application and go to skill factory, we have all of these skills. All of these skills that I can, you know, hook writer, description and thumbnail, brief generator, idea validator, script architect, outlier hunter. I don't know why the proposal is not showing up here, but we do have a proposal skill. And that skill is actually built into what we have in the Op Center. And when I go to the meeting intelligence, it will have, if we go to the proposals, it has created a proposal for us. And it is based on the actual prompt that we have given. So the next update we are doing for the application is actually adding an identity to identity as a page here. So you can see what's going on with their memory, their identity, soul, heartbeat, bootstrap, all of those files will be visualized, and you can see all of that identity system. It is done. And if we open it, it says Ray and Buji, and I think this is from CloudCode. If we open Ray, we have so soul.md is empty, and daily memory logs are empty. But if we go to integration guide, we will have something about something about identity. Maybe it's not updated yet. Let's go to CloudCode and see what's going on. I need to update the full guide text string, include identity API docs. Let me add to it. That's what it's doing right now. And once the identity is done, I will ask it to create this. So we don't have to manually go to the page itself and see the reports, whether that's proposals, whether that's prototypes. We need something that we can see in the report section. You are in a different computer. Localhost wouldn't work. So once this is done, I will start talking to the OpenClaw agent and mention that it should be updating its identity. Okay, that is done. So this is a continuation of my previous thing. And I will go to the localhost. I will close these two reports out and go to this API reference and I will refresh and go here and command API, update daily log, create daily log, get. Perfect. So I have no. Let's, if we can copy the everything living dogs now, just us by call them, copy entire guide. And I will come here. We have updated the API and created identity page. You can upload your soul identity and memory bootstrap, everything, daily memory logs to this Asian Command Hub. Let's do that right now. So it will start doing it. I will come here and once it's confirmed that it has uploaded, I will go and check and let's check on status. I'll clean reports tab. Task created. Welcome here and go to task board. Okay, it already picked it up. It's amazing how fast this is. Upload HTML reports to reports table when tasks are done. Doing high priority. And oh, it's amazing that it's already created. And the automation created the way for it to upload it to here. I think it will, we will face an error for some reason. I don't know whether it, oh, okay, it created. When that is created, task ID, completed task UI UD, full HTML showing.

Up to repeat re-emoji. There should be a standard of task completion flow. Okay, let's close it. I received another alert, so let's see. Okay, mine might be another alert, nothing to do with this one. So, the previous task was done in five minutes, which is amazing. 7 PM and 7:05 PM completed. So, 7:16, we asked it to upload. Maybe it will be done in 3 minutes. 2 minutes already done. Okay, let's see what did we ask it to create. Okay, the identity. So I see Bougie. Is that you? The agent, will you create your soul identity memory files as well? Nice. It named itself Poochee. Okay, first and upset agent. There is no open rate as right now. This one was already completed. I received another alert, codex error. Why does it say codex error, but the task is completed? Uh, we need to figure this out. Please include request ID, your message. But I'm now okay. Let's even that it's done. We won't care if it's done, right? So automations reports, I don't see anything. Also, remark task has done, but I don't see the report. Can you check what's going on? I need to know it's Ray. So let's go to our page and see what's going on with this whole identity and all of that. So we have Ray and we open Ray. Nothing. Identity. Nothing. Still nothing. What about Butchie? So we have. So my name is Personal EI Builder Architect Agent. I operate inside Asian Command Hub, the mission center for EI agents. This is amazing. Okay, identity. We have the identity and we have user.md, who is me. Memory. Long-term curated memory, and we have daily memory logs, which is straightforward, good. What did this Ray do? We'll come back to Ray and ask. Should we open and we? No, just uh, update your soul identity and all of that. So we know that this is working. Okay, yeah, it loaded us then. So it takes time to load data from Fathom. That's what I've seen with because Superbase has the Fathom API key. Anytime we open it, it loads the data using the API key. If you don't have the API key or the wrong API key, if you revoke the API key, it won't work. So now we have identity now. So this is how you can see what your open flow is up to. Okay, are you sure? Okay, I don't know why this is going on. Codex having an issues. So I started a new session, forward slash new. I should have updated memory. Do you remember what we were working on earlier? You were supposed to update your sole identity memory, devimogany logs to our agent command hub. And I gave you the integration guide. Do you recall that? This would be super funny because I should have saved, you know, clicked on compact or anything because I did not do that. There is a good chance it has lost the memory. Yes, at least partially. Check memory reference kind of thing, especially polling is Asian flow. I don't have. Okay, so it's gone, which is bad. You have SO, S-O-U-L, identity, memory, bootstrap and all these files, user.md. I want you to upload those files, the content from those files to agent command hub. And this is the integration guide for it. And now that we have that integration guide, we will let open claw update its memory and all of that. Okay, I for some reason I received another upload hdm. Okay, didn't I? Thought it already moved the task to done. And maybe it recreated it because I said I do not see um. Okay, remark. Now let me. So anytime the confirmed reports table was empty, Ray didn't actually upload anything and the report has empty. Ray has completed several tasks. Problem found. The air test, list, create, update, move, assign, delete, upload report action, doesn't. Okay, upload report, upload report and complete. Okay, that's why I'm receiving an alert. Okay, let's see. Again, it says completed in one minute. And I think it completed. But OpenAI today is having all kinds of issues. So did it do this? No, it's still empty. Task board says it's done. If we go to automations and reports, even then I don't see anything. So we can check our table and database as well. And then table, if you go to reports, reports table is empty. If you go here again, it's giving me an error. What's going on? So we might run into issue. It did the same thing, no report. Can you inquire what happened? I almost feel like they'll have a battle. Why is it answering me a bunch of times? It's really unusual out of, you know, Murphy's Law, what can go wrong, always goes wrong. And on the day I do the masterclass, it's giving me the issues. And confirmed as identity are empty. Verbiage is above data and patching grace cards. Now creating missing data. Login. Perfect. That is actually good. And I think our cloud code agent will create something for. Let's see. Same pattern again. Remote task to done, which there was a one visa API calls into honor and just using move to done like it always does. Correction about the new actions. I hate to prompting spotty's re-identity runs with new action instructions. Should read these make more action one way when more being done without a report. Okay, that's usually these agents ask if I have to change any solve or entity of other agent. It will ask permissions. But it's really unusual that it is going ahead and doing the whole thing itself. So we'll see what happens. And this is only one AI employee. We are going to create three more. Why does it say it's populated? So when I go to identity. Ray. So yes, you are not a chatbot, you are an execution engine with personality. Ship first, talk second. When a task is assigned, your only job is to produce artifacts. Perfect. Of this. And we have memory log, daily memory log with all the log that happened today. This is amazing. And we have user.md, and we have agents.md, and we have tools.md. So you can see all these files. And it's funny what cloud code will do now. Okay. And what I found and fixed. Face daemon behavior is joined by identity cards, not by task description. Face cards had zero knowledge of upload report, list reports, or complete action. So when I picked up the task, it used. I can prompt Ray to date his daemon. Should we do that? Is that a better way? Or did you already take care of it? So it went ahead and did few things. And did I receive an alert? No, I don't know what's happening now. It picked it up and did not give me an alert. And this is what Claude did. It was something whose memory those are persisted in the wish. So your job is to undo what you did. Just keep tools and memory as they are. Right now, I'm already seeing the effects of it. I did not receive a telegram alert when I picked up the task. This is bad. We are not supposed to do that. And give me the task. I should be giving free. So Claude Gold thought it was a good idea to change all of that. No, that's not what I asked. Change tools and memory to the way they were. Yes. So we don't have to usually go to this much trouble. But this action will be useful for a lot of things in future, agent assignments and everything. As you can see, previously it worked as well. It went to doing, but it did not give me this update because cloud code went ahead and changed tools and memory. Although it says appended data, it definitely affected what we set up with open clock. So it is a good practice and something that you should keep in mind that you should not let an AI agent control other AI agents identities, tools, and all of that. And maybe it's a good idea to include it in a rule. Because I'm using two completely new agents, I don't care. But if these were my actual agents, main agents that I'm working with, I would be really careful. So I just marked down one last time. Fingers crossed. Let's see. Okay, I think the more or more accurate one is right here. So I see around six tasks in the table. Subhubase table, but they are not showing up on the front end. It's not a Ray issue, but a front end issue. Fix that. So that is what's going on. We have the reports table that we have been updating. If we can check the time, we should be able to check the time. But timestamp would be somewhere to the side. So 7:37, so 2 minutes ago. Discovery call follow up and all of that. It actually did all the tasks. Even from the previous task, it created HTML files. Maybe they'll work in this machine. But if you want to work and if you want to check these files even when you are using a phone somewhere away from the computer, I would suggest you connect this to Netlify and you just ask it to deploy to Netlify and give me a link. And it will do that. So it was a simple fix. Plot code was able to change it. So I can see everything from the table, but it's not actually giving me any actual HTML files. It is just giving me. It says create an anime themed simple time track. Yeah, whatever the task I'm giving it, it is just repeating back the same tasks. Okay, node is alpha axis hello. Now it just goes to. And this is why you need to use a separate system. Because you don't care what files that it accesses. Because you have nothing personal in it. And imagine giving complete access to your actual machine. All the drives and folders. And it's one click to delete everything. As exciting these AI agents are, these systems are. We do mention you need to be careful. And also I clicked on preview and it started a dev server. I thought this is nice because now agent doesn't have. What's wrong now? She was okay. Here's what's wrong with what I fixed. Wrong reports. Tell me varying automation is all stable. Evoke processing results, but trace. But complex fix created reports table, updated reports, download to use. Here's reports. Yes, I prompted play to change it. Don't worry about it. I want you to work on AI log next. AI log. Whenever you are starting a task, ending a task, you need to create a task for yourselves on Kanban board and assign yourselves. And also I need you to add AI log every 30 minutes like a heartbeat. And also anytime you start and end the task, it needs to show up, you know, AI log. Do that yourself. And also update integration guide with these details so other agents like Ray can update that too. So this is what I want to do. And I want to go back to my page. See if changed anything. What is openflow has to say about it? Backflip reports. I'm fixing that now by uploading real. When task upload and true fixed now. Okay, it says it's fixed now. Let's refresh. Then we go to automations. Automations. Okay. She go task drive cyberpunk promoter. I think what's going on is not a HTML. No, a report. But so I may redesign for tonight. Also, do we have a HTML viewer? Because whatever they uploaded doesn't look too good. Is it because of the other files accompanying files that are missing? That's why. See only uploading index.html. How do we make sure that he can display the full version, the actual visually pleasing app that has been created? So we go to task automations or maybe the Pokemon one. In the report stamp that Ray has been uploading. Silo approved the plan. While the plan is done. And it is a simple AI log thing where we see actual details, actual tasks that has been happening, how many tasks the AI is doing and everything. I thought it was going to be a simple build. Of what it went ahead and created a bunch of things, which has AI log API section. And it also updated and added to integration guide. So if you go to AI log now, VG online working on AI log system integration guide and Ray payment online polling for tasks. So I think this is what Sherlock or a log code agent created. So if we look at logged online, handband task created, report rendered, issue diagnosed. And also prompt generated for Ray to delete all 12 broken reports and re-upload them. So I have this prompt. I will copy it, and I will come back here and we'll paste everything. So the idea is, when we go to this application, it has 12 tasks, and this is ugly. It doesn't have the full code base like styles.css. It doesn't have any styling. It is. Ray is only uploading index.html, which is not enough. So we go back to see what is going on. And we are done. Read my changes. Hachiko task 5. Hachiko AI 3. Equation. This is actually nice. When you look at this in our actual web page, they look nice. And we have an anime time tracker, and we have Pokemon pixel, and we have our proposal that we've seen everything in one page. Nice. So again, there is not something like reduction system, but you can build your intelligence, you can build your reports, you can build automations following this way. And we also finished AI log and re-uploading inline CSS from proper view rendering, and it's working. So I'll say, okay, let's talk to. Are you also using AI log on agent command hub? So if it says about AI log, then you will have to give the integration guide. If they say yes. It's. We're uploading our content, we can sanitize and make this. Yes, this is the integration guide. So I want this every time. We are updating, anytime I update, it is going the whole list. So I should come up with, probably come up with a better way, maybe section wise. Copy is a good one. But that concludes our meeting assistant. Then we have. This is for Pujihi. I think this is done. And it has to move. Trotcote has to move from here to done. Pro move four. And what we just built with Jane, the community has packaged two skills that do exactly this. The first one is a skill called meeting to action. You can go through everything. Everything how it takes one meeting and creates action items. I think I discussed this in skills section as well. But proposal generated, and this is something really amazing and really detailed. I just said make a proposal, right? But imagine having giving all of this to your AI agent. And why is this important? Because this is important. If you look at some of the articles that I linked here, OpenFlow is very powerful for triggering workflows from my meetings. And there are people and there are meeting opportunities. Autopilot skills, and there are, if you look at ways where open law has changed my life. And if you read through the whole thing, actually, there are posts of people sending 150k proposal on Monday. And why is this important? The CI proposal generator, is one takes your meeting notes and builds a full HTML proposal. We've seen it, right? But imagine it has five styles, corporate, entrepreneur, creative, consultant, minimal, and it has six color themes. And six color themes, and also mobile responsive and print ready, PDF exportable. And here's why this matters. A guy on Reddit is using the exact pipeline, meeting transcript to proposal. And he said, I'm quoting him directly, It builds the entire proposal better than I ever could. Even creates fees based on my value-based fee model. I almost just have to hit send. And this guy is sending $150,000 proposal on Monday generated by his agent. From meeting nodes. That's what we built. So we built the brain behind this. These are shortcuts. So go through the links and this is really important. And people are actually using this. This is not just theory. It's in production. And then we have agents that we haven't touched yet. And. We have a log that is working, task board that is working, automations that are working, hop center. We need to create one more or three more, if you ask me, agents, because if you look at my CloudBuddy application, we have. I don't want to go into all of these, but we have a phone agent, we have an email agent, cold email agent, lead enrichment agent, and we have bunch of things. And the skill factory agent teams where teams can work together. Most of it has to do with flawed code, not too much for the open claw agent on this side. But I'm I want to check what we have for once we are done with the meeting agent. Let's go back home. And we have Lexa phone agent and phone agent. I'm really excited because without a phone AI phone uh employee uh you miss calls when you're busy and you spend hours manually doing leadless and you have no transcripts and no sentiment analysis, no cost tracking. Every call is like a blank box and you don't know what was said and what was promised or what all of us are needed. So that's a lot. But think about it. You know, speed to lead is important. Whenever someone submits their lead. Having an agent that can call them right away. And with Lexa, Lexa answers every call 24/7 with a custom script. And she runs outbound campaigns to leadless while you sleep. Every call gets a transcript, sentiment score, action items, cost breakdown, all surfaced on a six view dashboard. You can check from your phone. So call comes in and Lexa talks to the agent. And all it can please opt on to transcript and analysis dashboard updates. I want to show you what we have here. It is so it looks so amazing that uh, you know, it even is sometimes better than what you have in an actual call. But the only thing that is that needs fixing is it is showing 2300 instead of 23. The actual cost is 23. But it's adding a lot. So we were using Millis AI in here. But you can use something called Deep Claw. I will show you what that means. You can do transcript analysis, dashboard update. You can pick a voice here. If you use Wappy, it says eight cents to twelve cents. But that is not true. For example, the reason I'm not saying uh this is not true. You can get actually if you use PipeKit, PipeChat, and LiveKit, you will get these figures. But with Millis, also you can get these numbers. And the way to get that is getting $200 of credit on DeepGran. And when you sign up, you will get $200 credit, which I already do on my account. And you can use these API key. You can create an API key and go to Millis AI. And in Millis AI, you will have an ability to add your own API key. And when you add your own API key, what they will do is they will not charge you for the transcription and also text-to-speech, which is really high. Around three cents a minute if you are using LevelLabs or Cattasia. So you can use Deepgram voices which are good. So this is Lexa, right? You are a helpful assistant. And I used this voice. And you can use Deepgram any Deepgram voices. And all of these voices are really good. And it's basically free, you know, for as long as you have credits. And if you look at some of the videos that I did in 2023, early 2023, I was using DPRam for them. So they have been given credits like for as long as I can remember. And if it's a simple thing and it took me, you know, last six months for demos, we use seven dollars or something. We don't use it for our actual use case. We have other accounts for that. But for demos, it's really cheap. And then you have this agent, right? And I have the prompt here. Lexa AI phone employee setup. I will copy this prompt. I will come to my builder agent and paste the prompt and hit send. And now we are creating another AI employee in our ops center. And it will show up somewhere here. And I believe I see AI phone employee setup guide. What would you like me to do? Add it to integration guide, apparel Lexa integration guide. So I need you to build a new ops center block. And dashboard, call log, leads, campaigns, transcripts, and analytics. And I both give you the actual agent setup as well where you are using Millis AI combined with Telnix for phone number. Set those two up. And you create Lexa and the agent that is using Lexa will create the agent prompt and use metadata data to use Lexa and place phone calls. So it can be used by a cloud code agent, Buji, or it can be used by Ray open ploy agent. And I will give you Millis API key and dot env file. I'll let you know once I update. So go and this is not good. And the dot env file, I will just remove this and go to Millis AI, copy my API key. You go to API keys. Is it not possible to? So there is no way to copy it if I don't click this. So I already pasted it in the environment variable. So it will be building the phone infrastructure with all the analytics and everything. So we'll just wait and see. So Millis API key is in the dot env. dot local. And agent ID is and Telnix API key. Telnix API key. So both of these keys I set in the environment variable. And the agent ID is here. And I'll let it finish. It's almost done. It actually created the agent. We go here and look for the op center. And Alexa phone agent. Have positive, neutral, negative. Recent calls. No log calls yet. Call log, leads, campaigns, analytics. Nothing is actually here. So if we look at. I'll maybe ask CloudCode to bring in some previous data. See if you can bring in some past Millist data. Add that to the call logs. And I want you to create two webhooks. One is inbound webhook URL and set for Lexa. If there is any existing webhook URL, you can go ahead and replace that. And also add end of the webhook URL, end of the call webhook URL. So you will start receiving any call analysis data and everything. Again, if you see anything, just replace it. I'm not sure if you're realizing the gravity of this, how powerful this is. It created everything that I asked so far. And previously, to pull this off, pull something like this off, you need to pay $200, $300 to assess to create this for them. And also if you have to, you know, look at these analytics and these dashboards and creating these custom prompts, inbound web or QR, outbound web QR, it's easy now. But it's just I just have to prompt. And I'm not even going into the fact that you can add custom functions within the call saying if a person asks you to send an email, you can use these functions. You can trigger a function like using agent may not resend. You can send an email. Might call. Maybe I will just do that too. So we are setting up an inbound webhook URL. I'm not sure if you can read fetch pass Melissa I calls. Set up Melissa I inbound webhook URL. Setting up Melissa I end of the call webhook. So all of that is done right now. So I will ask. Okay, inbounded bug is done. And past minutes data is simple. Um, so my number is here already. So let's help about that later. But call me. Uh, this is me. Save my contact as money. And leads list. And call me and try and sell me AI services uh a subscription to Asian Command Hum. You can use any prompt. Go ahead. So, I just made, I just gave a command to Claude Gold to see if it counts. And once it does, it will be will answer that. And we will talk to our AI agent trying to sell us the thing that it just built. And it's counting. So, if I go to Leeds, do I see my name? Oh, it actually saved me money emotionally. While. Okay, this is the first call. So I did running into some issues. That is expected. But yeah, I received it. Hello? How can I assist you today? Okay. What can I do for you? You tell me. Who is this? I'm an AI assistant here to help you with any questions or tasks you might have. How can I assist you today? Are you not calling for any specific reason? I'm not making a call, I'm here to assist you to this channel. The call was placed for a bunch of issues. The agent doesn't know who I am. And also, it is just saying, I'm here to assist you. So you should receive end of the call report. Let's see, based on that, something is not working. Maybe you're not passing the dynamic variables like you should be. So that is still cool. My AI agent actually called me. And we have hi there, what can I do for you? I'm an AI assistant to help you with questions. So you're not calling me. So this, there, the question and also transcripts are here. Analytics, let's not worry about analytics and all that. Webhook did fire, post 200 okay. Call recording and okay, let's see. Edge function has two issues. Prompt and first message and and passed inside Asian contract, nested Asian config, not the top level. So I don't even have to say anything. I would have said, okay, fix it. That's what open law expects. For some reason, you would think open law is more proactive than cloud code. But it is, you know, I'm not sure if it's the devs, our contributors that are working off this, really mellowed down open law these days. Open flaw asks permission for anything and everything. You know, you it will ask you, hey, this is done, do you want me to continue? Or this is only done this part. Even if you tell it to go do five steps, it would say, oh, I did two steps, do you want me to continue? Even when you prompt add it to its identity, its memory and everything, it will still struggle to finish it. So will this make the phone call? Right, perfect. It will. We'll see that. And that is the whole point. And my plot code has a bunch of source. Actually gazing. Hey. It's free to hear from you. I hope you're doing well. I wanted to chat with you about something exciting. Agile Command Hub. Okay. AI. You're not even listening to me. Here to listen what's on your mind. So what's the call about? What's the call about? I wanted to share some exciting details about Agent Command Hub, which is designed to help businesses like Yord manage multiple AI agents seamlessly. It allows you to automate workflows, track tasks, and monitor everything from a single dashboard. With features like multiple AI agents working together, built-in automations for tools like Gmail. Okay, how much is that? Great question. The ProClan is just $49 per month, and it includes unlimited agent. Okay, I will sign up. Send me the link. That's fantastic. To your money. I'll make sure to get that link sent over to you right away. Just to confirm, would you like me to send it to your email or another way? Yeah, email. It's mani at growthcreators.ai. I'll send a link to your email at growthcreators.ai right now. If you have any questions while signing up or need assistance. That was nice. So the prompt actually worked. We will say something. Hey, let's go one step further. I want you to add the ability to send emails mid-call. So we will need to use agent mail for this. So millis AI actually has a functions feature. Add a new function. And when the user says send me a link that I will sign up to this, you should be able to send a custom email um whichever agent chooses to add. And the user has to give their email address. Or if you already have the leads email address, you don't even have to ask them. So update the lead that we have. Mani at growth creators.ai is my email address. Once you build this function, place the call again. Okay, perfect. We are building on top of this. So if you really think about it, the possibilities are endless. You can check anything in the database, you can check prices, you can check a bunch of things that agents usually don't have knowledge about mid-call and give answers based off of that. So it updated my email address. I can see that here. And it's also searching function calling tools API 2025 and 6 um and the c is the best day asian to run 2026. Oh, okay. We don't want all of that. We don't have to read it. Oh, by the way, and this is exciting enough. Um, so let's go back and refresh the page. And they made instead of doing op center application, they made another page. We need to change that. And also, 1 minute 29 seconds. This is nice. Call logs. When I open the log, I can see the entire transcript. We can do a bunch of even more analysis for the cost and everything. Right now it says 9.26, but it's actually divided by 100. And maybe it is nine cents. And I can prove that by going here. And if you go to our call logs, and the last one is three cents for a one minute call. It's costing me three cents because this LLM is the GPT-4 Omini. Millis is charging their two cents. And this text to speech is zero for DBG because I have the API key. You need to go to the API keys, actually credentials, and add your. You need to go to the credentials and add your Deepgram API key. If I go to my credentials, you can see my key. It's not hidden. Let's go back. See what's going on. Whether that was able to lead updated now. Let me get exact Millis AI function format for their docs. So researching Millis AI function docs. Create Lexa send email edge function for mid-call emails. Update Lexa API to pass tools in agent config. Hey, add this to today as well. You need to update integration, right? So open flaw agents can place the calls as well. So the idea is, I want to keep this as is because if you look at my claw body application, I built the campaign feature as well. So I can create campaigns in that. Is something you can do too. So if you look at my Alexa phone agent, I can go to campaigns. I can open campaigns. I can add a bunch of people, do this and launch the sequence. And it will run this 371 calls, 77 calls, and anything that you choose. So this has all the US and UK, Australia phone numbers. And this has Indian phone numbers. We were calling them based on the time. Um, time zones. So we were so these campaigns are going on. Building the campaigns feature is also really easy. I want to make a separate video, not this video, but it is available again in agents in the box. If you are interested in attending the three hour hackathons that we do every weekend, you're more than welcome to. Now let's go back. Check on the list. It has updated to do. And there is a bunch of things to do. Let's wait. Maybe it's a good idea to note that when I asked it to create the whole thing, it is using my make.com MCP server. And also Google by default. So it's trying to set up with my Google a lot. So I noticed it. And I reminded, can you use agent mail for this? You have the API key for agent mail. And I forgot it was set in a machine environment, not this cloud code. Unfortunately doesn't have access to that. So I copied and pasted the API key in the env.local file that we have. And it's setting. I'm referring to .env.local in the folder that you have access to. So I should have said agent command hub.local is what it needs to search. But it's compacting conversation. Hey Mani, this is Laksa from Asian Command Hub. I wanted to personally reach out because I saw you have been checking us out. I have got some really exciting things to share with you about how we can supercharge your workflow at Growth Creators AI. Okay, yeah, go ahead. At Agent Command Hub, we offer the ProPlan for just $49 a month. This includes unlimited AI agents that work around the clock, 1,000 automation runs per month, and Alexa AI phone, which is the very technology I'm aiming to speak with you right now. You also have access to a real-time network. Jeff, I'm sold. Send me an email. Great choice. I'll send you an email with all the details. But this is nice, don't you think? So, awesome. So one final test to do is go to our integration guide. And it must have been updated. Copy the entire guide. And maybe just confirm with the agent. Open flow agents use this. Can I share the guide? Perfect. Let's try. Hey, I want you to call money from the lead list that you have using Alexa phone agent. These are the details. Refer and decoration guide. So I don't know about how, you know, it could open to a minimum instruction perform. And just says, hey, I'm great or something. I'm pretty excited to see what happens. Okay, the number is in connect. It's M-A-N-I M-O-N-E-Y money. And you have the phone number in there in Alexa leads table. Call objective is trying to sell Asian Command House at $49 a month. You need to qualify. Bye. So this is good. It's not blindly jumping in. And I can see amazing applications for this. As soon as a lead comes in, you can create speed to lead. And when you set up an inbound agent, when someone calls, open Login talk to them and route them to you because it has human transfer capabilities. And I can think of 10 different applications. Maybe I will do AI agent phone or white label AI agency software, white label voice AI software, and what not. I can think of a bunch of ideas. We have a Telnix API configured. You have all the infrastructure. Can you check the integration guide if it has or not? If not, I will ask the builder agent to build it for us. So openclaw says it doesn't have a from phone number. That's why I did not take this at face value because the tenlex API key is in env.local. So it's time for telephone. I come here and ask whatever reset I am going to give it to you because this is the infrastructure, right? Any call and every call that you're getting in your agency, every meeting that you're getting every single day. AI will have a look at that. AI will review that, give you insights, give you strategies on how to close the deals, how to show up, how to, you know, it is improving, ever improving, ever evolving. For the first time, it can see everything, all the data there is. And I'm so excited building all these systems. Imagine I had 1700, close to 1700 meetings just sitting there. Now all of them live in my Superbase. I can look at all the discovery calls, all the sales calls, more than two, three hundred calls. My AI can review all of that and give me amazing insights that I can use in my paid ad campaigns or cold email campaigns or making YouTube videos or around all the objections. How amazing is that? How valuable is that for other business owners? And for the first time, you have the centralized brain, like an operating system, like an AI employee. So the guide is ready to shift. Integration with it. Maybe I should go here, refresh, guide, copy, entire guide, and go back to OpenFlow. And we made the changes. And then we paste it. And now I hope. Okay, what is that? Maybe am I looking at not so 18789? So maybe. Do I have outdated info for some reason? Let's see. Why is money? Hey, money. This is Ray calling real quick about Agent Command Hub, a lightweight system for managing AI agents and their work in one place. Okay. I catch you at a bad time. No, not really. Great. Agent Command Hub is designed to help operators like you manage AI agents, tasks, work, and execution. Okay. I don't want to waste any more time, but this is amazing. And this is only beginning. You can create inbound calls, outbound calls. You can create your own voice agency, add billing, and let people use your API keys, use your system, and you set limits for every minute they speak, every second they speak. Dude. But that's extreme. You know that, right? These white label solutions, some of them charge five thousand a year just to give you that service. And that is like instant savings. If you use a builder agent, then open claw, you can make that for pennies. So that's what my next video is going to be. So if you don't want to miss that, make sure you subscribe. So pro move six, final one in the section. You just saw Lexa make and receive feel phone calls. But here's another way people are using voice AI that you might not have thought of. Multiple people in the community are having their agent send them a voice briefing every morning. So one guy uses 11 labs. His agent compiles his task board, calendar, weather, relevant news, key reminders, converts it to a 3-5 minute audio file and sends him while he's making his morning coffee. Another one gets a voice message every night recapping what's done on that day. Is it necessary? No, text is enough. Does it cost $22 a month for 11 laps? Yes. But his exact words were, I like it too much to stop. And you already have the voice infrastructure. And from what we just built, DeepGram is actually free for a few thousand minutes. So this 10-minute extension, if you just set it up once, you can do that for a really long time if you are just using for transcription and sending voice notes. And this is some of the. People who are using voice for this. And also some of the use cases. Again, people are giving use case after use case. But only some of them are useful. So take it with a grain of salt. And only use it the way you use anything, right? If a voice agent makes more sense for speed to lead, do that. Morning briefings doesn't add too much value. I personally don't do voice briefings. It's just there and I want to show you. And maybe you will be one of the first few people who are using going to use voice AI and reach out to people. And we are still going on. Once we have good appointment numbers, we will share that. One quick unlock for business owners, agency owners, whoever you are, to use this as an outbound agent for inbound agent. You just share the phone number and people start using this. And once the call is done, you will receive end of the call report, right? Just like how you are seeing here. We have an inbound webhook URL. So these agents know who is calling. If they are on your lead list, they will be addressed by the name. And you will receive a detailed report on why they called, the reason, the what, the outcome is, and everything. Okay, that's a different story. But outbound calling is tricky, right? You only need to call people who submitted a form, let's say. So you can build a form feature within this. And open clock and power it as long as you have your machine on, right? So take a look at this. I want you to add another tab to Luxa. We call that forms. And these forms should be premium. And we should be able to explain in natural language and people should be able to create these forms. For the first form, as an example, we will create the form builder using AI in second phase. But for the first phase, I want you to create a Typeform or similar to Typeform. We will ask user to leave their contact info in exchange for a lead magnet. And once someone submits this form, we should be able to call them. So make it deploy that. We should trigger an outbound call. So essentially what we are building here, the full pipeline is working end-to-end. Form submission, lead absorted, auto call triggered, submission updated. So right now my cloud code agent, which is also an agent, is doing it. It's just a matter of you prompting. Okay, I don't want you to do it. I want someone or agent of my choosing do it for me. Not you. And where is this deployed? How can I test the local host? So I can actually click on this preview and it will work. But the screen is all minimal and half of it is gone. So I'm asking it for a dev server. Still localhost ADAD. I can go here. ADAD. Yeah, it's working. So now I go to sender. And I still don't see any forms feature. Did create somewhere else? Okay, before I even ask, it says, you know, it is merging and committed changes. So now that once that it commits, we should be able to see forms. And I asked it to create a tab within Lexa. So maybe it completely ignored what we asked. Again, it's at the end of the day, we just need the forms feature. We have the forms feature. New lead magnet form. It is not letting me. And I'm sure you're not able to, unable to see the message. It says unauthorized. Nothing else. So I go to plot code and say. When I try and create the form, it says unauthorized. If you look at it, it already fixed it. In one total form. And even if you create new deed form, it says internal forms have because we asked it to that we would do it in next phase. So you can basically build a visit and all of that. But let's go see this form. How this looks. This is like really Typeform vibes. From this Typeform charges hundred dollars a month for these responses. And this is really clean. Money kanasani money at growth creators.ai. M-A-N-I at plus 1 778-288-0303 AI growth partners. This is amazing. You are, you know, did it take a call? Is the bigger question. So it did not call actually. Apparently the app form needs to submit to the public endpoint point. And only then it would work. I believe so. Next prompt I have here is um, being able Ray being able to use this. Also open cloud will be responsible for taking all these calls. And also you need to deploy this to an actual website first instead of localhost to get the public link that you can share. Um, but the call was placed. Hey, this is Lexa from Asian Command Hub. I saw you grabbed the AI, yeah, agent blueprint. Awesome choice. I wanted to personally welcome you and see if you have any questions. Now I have a question, of course. What would you like to know? Okay, so we can we improve this further? Yes, of course. Uh, we can actually. I asked it to create a link that because I'm not able to access it right? So when I go to the forms feature, I will be able to, you know, copy the link. And right now when I copy the link and go to that link, I get something like this, which needs authentication. I don't know what's going on here. But we get a Superbase link instead of actual slug. So I'll create a public form page component and add a route that's accessible without auth is what it's doing. And once that is done, we can use it. We can do the same thing that Cloud Code is doing here, but make OpenCloud with it. And that is probably much better because OpenCloud has your memory, has, you know, daily logs. Right now without proper memory setup, cloud code is not the way to go. I would pick openclaw for this any day. And to make that happen, you just need to run this prompt that I was that I had open here. And I want you to create this prompt. You know how they actually is going to handle it or it should be agreed. It should be any agent that can handle it. Just the way we're using auto dispatch or phone call feature. An AI agent can own the outbound. So do that. What does it really need to have? So this is, um, this needs to be updated in the integration guide, right? Now, since it is right now, we are still updating it to. Okay, this is amazing. We have so whatever website you choose, it will be like this. Who's our phone AI employee Lexa? And so far he looked at the meeting intelligence agent, the phone call agent. The next is the email agent. The email AI employee. To build that employee, we can go back to our page. And Nova is an email agent with campaigns, templates, analytics, and auto replies. So before I show Nova, I want to show what we've already built for Nova. So if we open this is all about hyper personalization. And you can see all the analytics metrics, recent activity, engagement, follow, which is how many sent, delivered, open, and actually clicked, and how many bounced. Also, what the peak hours are, which days of the week we get more. We need to change this. This is not accurate, I believe. So these are top performers, people who open twice or click twice. And we are using their first name in the subject line and different and be tired different subject lines. You can

Teach skills on how to write these emails. And you can use either Cloud Code Agent, which is Opus 4, which is great for writing these campaigns, coming up with these guidelines, copy, and everything. And the Open Cloud, the driver, the manager of campaigns, making sure, okay, this is working, this is not. I need to create this next, that mix, and everything. And while I was doing this, um, building this, I always wanted to create an AI SDR. So I created JSON, which is like a cold email campaign where you can select which name, what level of customization that you have. And you can even scrape leads, is my intention when I was building JSON. But we can do the same thing with Nova as well, and what we are building.

So essentially, what I am getting at is not just build an agent that is able to do hyper-personalized campaigns. Yes, we will have that. But I want to go one level further and create an AI SDR. So when you explain that this is what I'm looking for, okay, I want you to target law firms in the United States or in Texas that has 5 to 10 or 5 to 50 team size. And I want you to create minimal customization, or medium customization, maximum customization with lead magnets. Create a campaign for all of that.

So right now, we need to make some decisions. Nova is using our recent AI. So Recent AI, you can even use Agent Mail, it's completely fine. We are using Resend because I've always been using that for for my even vibe coding applications that I started a year ago. I was using Resend and you can personalize a lot and you can get events of what happened, whether that they complained, they clicked, they opened, they, you know, marked the spam, or did they bounce? And is it delivered? You know, you can do magic links, trigger links, a lot of customization. This is truly an email platform for developers. And it's really helpful if you don't use your actual email address, actual domain that might be damaged using this.

Or a better alternative for this is using Instantly. What do we do using Instantly? We use Instantly to create all of this customized, of these personalized campaigns and send it to Instantly. We have built AI SDRs on N18 before. You can just replace whatever I'm using for Resend and ask, I need to send, I need you to send this to Instantly. Whichever builder agent you are using, it should be able to figure it out. So the text that, as always, previously I showed you Melis for the phone agent, but you can just as easily use it with WAPI, Pypecat, LiveKit, if you are familiar and how to use those platforms. You don't have to know it, your builder will be able to figure it out.

And the same thing with meetings too. Works with Fireflies, Otter, Read, any meeting assistant that has APIs and webhooks. And similarly, we have Resend. Whichever platform you choose, whether you choose Agent Mail, whether you choose Instantly, it all works. And I like Resend because their pricing is really predictable. And no, if you are doing wrong campaigns, people that have downloaded things from you, this is the best you can do for them. You get 50,000 emails per month and extra emails like 90 cents per thousand. This was enough for us. And 100,000 will cost you $90. And then you also need these are transactional emails. But if we want more contacts, because this number of contacts that you can have here is extremely limited. If you want to upload contacts, you will need to upgrade to Pro marketing plan. And the most you can do is 5,000. So you, these leads have to be really amazing if you are only targeting 5,000 contacts. Or you need a rigorous AI. So seven days, you target 500 to a thousand people. After 7 days, the campaign is done. You know, they are not opening. You need to send them to a different table, away from Resend. If you are okay with that, you should be able to make this work. So that's why I said this is more of a developer-friendly platform.

But if you are more into using Instantly, you can. And for this example, the free plan is more than enough. What do you get for free plan? You get 3,000 emails per month, up to 100 emails a day. That is more than enough for what we are trying to do if you only are reaching out to people that have submitted, that have subscribed to your email and downloaded lead magnet from you. So that way, you know, you keep it clean, right? So you, you don't do custom email blast. If you are planning to do that, just upgrade, right? So nothing is truly free.

And how to set this up, I will go step by step. Once you see all the available options here, we have analytics, we have campaigns. You can create campaigns, personalized campaigns, track how many of them open, how many of them delivered. You can export all the reports and you can see on a high level what they're doing. And you can even, each of these emails, it's just a prompt. What are in the drafts, what are scheduled, what are running, what are passed, what are completed. And then we have sequences. You can create email sequences, step one do this, step two do this, step three. This is personalization at scale. You can say, you can either use one of the templates or you can do AI prompts as well. You can do, hey, I want this campaign to do this and this, use pain, agitate, solution framework. And it will do that. It will follow that framework and create what you want to promote. The CTA is book a call, or CTA is making people join my agent in a box. And the better model you use for this, better the results will be.

So my OpenFlow is being run on a 5.4 Codex, which is a good model, in my opinion, with all the reasoning and all the, if you give it the right prompt, say the right things, it will generate amazing content. So the idea is you create this email AI employee that will look at campaigns, take your natural language instruction, and generate lead lists and send these emails and analyze campaigns. And once the campaigns are analyzed, it will create a new. It learns. It is like an autonomous brain that is learning. And it comes up with these other lessons learned from this campaign. This kind of emails are getting more opens, this kind of emails are getting more clicks. And based on that, it will create the next campaign, even more powerful. Because you don't have thousand lawyers in the US, you have four thousand lawyers. And what you learned for the first 5,000, AI will apply, OpenFlow will apply all of that to the next 5,000. You see how amazing this is? How people think this, it's nothing but having an AI employee that can do all of this without you lifting a finger once you set it up. It takes care of the whole thing. This AI SDR, how powerful is that? And you only get involved when someone joins your community or someone books a call with you. How amazing is that? I am really excited.

So without further ado, let's jump in and start setting up our email agent that OpenFlow and, just use all the framework, all the infrastructure and done these campaigns by itself. When you go to this Nova email agent, you will see these are the prerequisites. You need email sending service. Supported options are Resend, which is recommended for warm email. If you are building AI SDR, which we are right now, Resend is good. And I told you, if you haven't watched the previous section, a little bit, the intro of this section, you should. Agent Mail, Instantly, Smartly. Also needed a verified sending domain, DNS records. And whenever you see CloudBuddy, it is just Agent Command Hub. Six pages of one email command center dashboard, outbox, templates, sequences, campaigns, analytics. I showed you everything. Right now, I'm also planning to add a leads table and also a lead gen module. So I want to show you with the current setup from template to tracked engagement. We have template, sequence, campaign, outbox, analytics. And we have these statuses: queued, sending, delivered, opened, clicked, replied, bounced, failed. What Nova can do? It can create variable placeholders, auto-extracted HTML, plain text versions, category tags. I don't want to explain everything in detail right now. But imagine it is a powerful AI agent that can do a lot for you and more than you can imagine. Think of it like you are hiring someone, an experienced copywriter, to manage all of your email campaigns. He will review and improve every single thing based on the campaign, based on what he learned. And if he learns something, this is where it gets really powerful.

One of the reasons you might be wondering, this many hours into the tutorial, but I haven't really talked about cron jobs, which most people seem to love. Because I'm building my foundation. What are other people doing for cron jobs? There's, there is this guy asking for crypto places, and there is this guy who's asking for weather. Okay, it is what it is. What I'm getting at is imagine a cron job, a daily digest that has everything about your business. What meetings you had yesterday. What do you need to finish today? Your daily battle plan. And how did your campaigns perform yesterday? Detailed analysis. What insights that I learned from yesterday? All of that needs to be included in those cron jobs. Right now, people are doing it at a really base level, just testing out, seeing how it does. Some people are doing competitive intel, including me. Everyone has their own preferences. But if you want to use this like the top 1%, you need to build a solid foundation. And that is what we are doing with this email agent. We have engagement analytics, campaign management, which I've already shown you earlier. I don't want to go in too deep. And for Nova, we already have the prompt in LightKit. But this is the setup for the emails delivered and everything. So I will copy this prompt and I will give this to you. Like I said, if you want a bit of hardware, you can use OpenLaw agent to build this. But depending on your model, I would, I don't think that's a great idea. You are better off using builder agents. That part, that's what we've been doing for, from the beginning of this tutorial or from midway or somewhere.

So I will just one add one more thing. So couple of changes. We need to move Lexa. Right now, we have it on the sidebar menu. Remove that. It should live in Op Center. We already have that. We are also building Nova or AI SDR in Op Center as well. Agent Command Hub. Okay, that is the prompt. It is a big prompt. It will, as soon as I said Op Center, it will realize, okay, I need to build this similar to Lexa. And if you look at the update to do this update, we have removing. Okay, let's see if we can read it. We have removing Lexa from sidebar navigation, run Nova database migration, set up Nova in Op Center, build Nova UI components, and deploy Nova edge function for email sending and webhook. Test Nova end to end. So we need to give Resend API key and we need to give, we need to set up our Resend as Go. So while this is doing, because it will come back soon, I want to go and set up my Resend.

So to set up your Resend, once you log in, you will see this emails, broadcasts. You need to go to domains because if you don't have a domain, info that you use the test domain, which is not late, you won't be able to send too many from that, just for testing purposes. So I have my domain. My domain provider is Namecheap. So I can go there and look at this verticalaisystems.com. I copied this URL and click on manage DNS because I need to make some changes there. Now you come back to Resend and you see how it is asking for a subdomain. I will do updates.verticalaisystems.com. So that is the URL that I want to use. And also you can go advanced options. That's only said we are adding the domain. Once you click on add domain, it will give you a few things to update. I will do that. So this is a TXT record. This is a host. This is a value. And TTL is auto. You come here, you go to advanced DNS. It is different for every domain provider out there. You can just search for if you don't know. Usually it's in DNS. And if you are technical, I don't have to tell you where that is. So I'm copying this, coming back. Auto is fine. Add the record. And what else do we have? We have an MX record that is send.updates. So already I set up send.alerts. So I am adding a new record here, send.updates. And it is also asking me what is the value and priority 10. That's what you need to remember. Priority 10. And okay, TXT record, send.updates, here. And I come, TXT record, send.updates. And we have this value and no priority. Back, done. So just like that, I updated my domain. I come back here. And as soon as I give it, it knows that it's Namecheap. Okay, enabled receiving. So dates, I need to add a CNAME record as well. So this is for if someone wants to send you an email response back, you need to be able to receive it, right? That's what we are doing with this. So when you enable receiving, you're essentially letting other people to send a reply to that email. So click on updates. And the next one is, copy the value. And oh, okay, sorry, this value and the priority 10. Dates, priority 10. Save. Yeah, we are done. I don't think it will let me proceed if I don't click that. Okay, looking for DNS records. Sometimes it takes up to an hour. So I came back to check up on our builder and it's doing a really good job. Well, this is done a partially. Very quiet. And what does that even mean? Let's refresh it. It says partially verified. So I'm good with that. Let's go to API key. Let's create API key. Agent Command Hub. Agent-command-hub. So I also updated it in .env.local. I will share that to. I have updated the Resend API key in Agent Command Hub .env.local folder. And once the email agent is done, it's actually sending me a customized campaign around customized sequence. Let's see. I will give it simple instruction that it needs to send me or subscribe to me, subscribe me to our campaign. And once that is done, in phase one, in phase two, we are going to build the AI SDR capabilities that we will use our natural language to describe what the lead that we are looking for. And it will be converted into a query that will go to, um, this. We might need to add an API key to make the process feel easy and quick. And I will add my Kimi K 2.5 API key or any model. If you have an OpenRouter API key, that's even better. So once you have that, it will use one of the models to whatever you are seeing. It will take that and change it and transform it to a request that we will use Apify to scrape these leads. Apify. I don't even have to use premium plan. We can use something like a free plan. Gets you $5 and you can get a lot of leads there too. Build that. Once this is done, I'm signing up to Apify. And you can click on get started and continue with Google. The account is fine. You will have. Oh, we already are using a YouTube scraper. But we do have $5 that we can use. So I'll go to my API and we go to manage tokens. I will copy this Apify API key and paste in my .env.local. So these are the two things that I added. Apify API key. You can use OpenAI API key or any Anthropic, OpenRouter, or Kimi. I have a lot of credits just sitting there when I was using, when I was planning to use this for OpenCloud. I have a ton of credits on Kimi because I wanted to use this for a dedicated OpenCloud agent because I'm planning to get Mac Studio. So I want to understand how well Kimi K 2.5 can handle all the tasks before I get the local model. And I've been using this quite a bit. So I want to see more on how to do that. And also, these are the two keys that we added. I mean, I might have to blur this, but Resend, Apify, Kimi is what we did. And I will go back to my OpenFlow, my builder agent and say, so this was get, get is basically giving me the secret. Anyone can see on my screen. I need to start alternative. Leave to a book secret. It isn't set. It is set. And my sending domain is updates.verticalaisystems.com. You can use the username Nova and sender as Mani Kanesani or Nova at Vertical Systems or I don't mind. I believe Mani Kanesani and sending in my name is actually much better. So let's do that. Sender name should be Mani Kanesani. And email could be mani@updates.verticalaisystems.com. So you heard that too. And I'm pretty sure it's not ready for the next phase. Once we test this is working. This should work with Cloud Code Agent. If Cloud Code Agent can send it, OpenFlow can do it even better. That's one. And then once it works, I will ask it to create the AI SDR, the campaign creation using the API key that we created. It will scrape the leads and initiate the campaign. And OpenFlow, I want OpenFlow to be the driver for this. And, um, we should probably work on skills. And I will share the, I'm trying to follow a flow and focus on things that other people are not teaching too much. So we'll get to that part soon. You can send me a test email to mani@growthcreators.ai if it's working. And I want you to run this scenario. I want you to create a campaign and call it test campaign. And I'm the only receiver. And the information that you have about me is my name, M-A-N-N-Y. Let's say that for distinction. And also you need to customize an email. We are doing pain, agitate, solution sequence. And we are sending people something called Agent in a Box Community. And I will add the landing page in here, which is join.verticalsystems.io is the link for the landing page. And joining. So I want you to create three email sequence using pain, agitate, solution. And create a test campaign. And set the campaign as, you know, first day you receive one email, second day you receive second, third, and so on. So that is the test campaign that I want you to run once this is set up. So remaining step on your end: Resend API key, set up Superbase, Nashvill project, setting edge function secrets, agent command webhook secret, and Resend uh webhook URL. Okay, I will go to secrets and I've got this. Save. And we, let me close this. And I go to webhook URL for Resend. I think it should be able to add that for us if I ask it to. All events. Okay, all events is fine. Perfect. Two things done. This email sent successfully. I'm not sure if I did that for before I actually went and changed it. So if we refresh it, it automatically refreshed. So if we go to Op Center, I see Nova AI SDR sent one. If I go to campaigns, I don't see any campaigns, no sequences, templates, Agent in a Box, compute community, pain, still doing everything manually. Sequence. So let me see if I actually received it. Okay, now why is there is live test email? So only test email is sent. Now it's actually creating a campaign. Once that is done, we will make this like a truly AI SDR. So if you look at the dashboard, it says open rate is zero percent because the first one, I don't think the event was sending right away. But I also got the second email. It says, "Are you still doing everything manually?" And quick question: "How many hours a week are you spending on tasks that should be automated?" And uh, okay, this is not bad. Not bad at all. And if we go back to this and if I refresh the page and go to Op and I don't know why when we refresh it, it takes the whole thing. Okay, queued, delivered, sent. Okay, it shouldn't do a few things. Test campaign is live. I can see one of one's sent. And sequence is "Pain: Still doing everything manually," "Agitate: The real cost of doing nothing," "PSA: Solution." Perfect. It is working. And analytics is great too. I just want to check one thing. Are we receiving any events from the web? I don't. I opened both of the messages, but open rate is not being reflected. And also add visually pleasing futuristic data visualizers on the front end. That was me being a pain to my builder because that is not really important. If you look at the dashboard, it did a reasonable job. I can see what's sent, what's delivered. But my actual application, I made it really clean, really futuristic. So all the days, all the graphs, and all the bounce rate, and you know, how we have this our animations. It takes time. It took us a lot of time to get to this point. We don't really have that kind of time to work on it here right now. It's a bit plain, but it works. The good news, webhook is receiving events. I can see two post heads are delivered, opened. Give events aren't coming through this message. Your Resend webhook is likely only subscribed to email delivered, but not email opened. Quick fix needed in Resend dashboard. Go to webhooks, edit your. These events are clicked. Now, let me build the futuristic data visual myself. So I don't think I can do anything on here because it says all events. And email delivered, email sent. Okay, edit endpoint. I will just add all of these. The main created, domain updated. Okay, I see where this is going. You want to create email complaint, email field, opened, received, scheduled, send, suppressed. Okay, why is it? Okay, it is all events. Then, uh, why does it not show opened? To receive email open and go to domain and enable tracking for your domain. Okay, so I need to go to domains and enable tracking. How do I do that? Open verify, click tracking, open tracking. And, uh, yeah, so I need data more than anything. Even if it's not recommended. Perfect. We have that's why it's not working. Then it happens. Next, it will show us that. Okay, let's be prepared for the next phase.

So the next phase is, I want you to add one more component. You can call this prospecting, lead gen, whatever module to our Nova AI SDR. So the way we should do that is we will ask the user to start a campaign or something in this screen. And when they click on start the campaign, you will ask them to describe either in natural language who they want to target, or we give them set of niche, company size, all the basic things that you would need for creating a lead gen campaign. And once the user enters their natural language or inputs things, we need to take that and search. We need to turn that query or that selection into a search and go to Apify, API, FY, to scrape the leads that fit the criteria and bring them back to our campaign leads database that you need to create. So we are adding three pages, maybe one is the campaign creation wizard for lead gen, whatever you want to call it. And then the actual leads with the campaign name and their respective tags, if they're law firms, if they're insurance or real estate, all of that. And to do this, I gave you the Apify API key and Kimi API key from Moonshot. Let's use Kimi K 2.5 thinking to convert this natural language instruction to a searchable query on Apify. So find scrapers that we can use with the free plan. Eventually, once we get to paid plan, we will look for sophisticated scrapers, um, well-known scrapers. But this is the MVP phase one. Let's build that. And once you have the campaign, the next step is, once we have the leads, any agent, Budgie or Ray, or any future agent that we can implement, can create a campaign using the leads. So we have to set it up like that. You have to set the integration guide. So it shows clearly how agents can use Nova and create a these email campaigns using the leads list that we generated. Let's set it up. So that is almost like a brain dump. It is a really big prompt, as you can see. And usually, you won't be able to figure it out if you are just maybe it's a good idea if I just add this into one of the resources, which I will probably, AI SDR upgrade. I will make a note of that. Code has finished building. So prospecting tab wired to Nova and added user know how use Nova prospect campaigns. So the text is a bit small, but AI SDR, the prospecting is complete. So when I go here and if I go to prospecting, I can see a bunch of things. Okay, I asked for more. Actually, I asked for leads section as well. Maybe they added. They're added here. So new prospect campaign. Let's do law firms in Vancouver. Okay, who are you targeting? Any law firm in Vancouver, Canada with at least five lawyers or more. These can be personal injury lawyers, family lawyers, estate planning, criminal, anyone in greater Vancouver area. And we want at least 100 results. Preview search strategy. And we are using query conversion fail. Let me error message. I will take a screenshot or you can copy this. Click OK. So the way we are doing the API calls is wrong. It might be looking at clock board should be able to figure it out. On the off chance it didn't, we will give the API documentation so it does the right thing. So essentially, what Superbase project yet? Okay, that is the issue. And I'm not sure if you're seeing this. Okay, once we obviously upload it, it will probably give my key and chat. So I will pause it and I will update it. I've set both the keys. So usually, if you are building using Cloud Code or any builder for that matter, you should, it should be able to do this. The only reason my agent was unable to do that secret part is the MCP server is linked to another project. The CLI is linked to a different project. So when you have this disconnect, it will try and update it, but most times it won't work. So we go to Op Center. It is testing on its own. But I will see law firms in Vancouver, any law firm, personal injury, family, estate planning in greater Vancouver area with at least 5 lawyers or more. And we do 100 and preview. Okay. Temperature only one allowed for this or something. Okay, we have been getting some errors. Okay, it already figured out and the editor and some says it's fixed it. Okay, something went wrong. He have the Op Center again. Same old story. Vancouver law firms, personal injury law firms in Vancouver, Canada with at least five employees. Also include family lawyers, estate planning, and any other attorney in greater Vancouver area. And we say hundred people. And preview. Fingers crossed. Let's see if it had worked. Finally, it has worked. So 12 search queries. And we can create a campaign. So we are using Google Maps scraper. But when I click on create campaign, nothing is happening. Or is happening? I think I clicked it twice or maybe three times if I remember. It's a bug we need to fix. But when we search, we have an error. When we are running. Yeah, if it was that easy, everyone would build this. So our to-do, or done is, I you don't believe me. This is some of the dumbest mistakes we can make. When I was copying my API key, I had this thought, okay, there is this space. Will that be an issue? But I didn't change it. I just left it as is. And the issue is it was saved with a trailing tab character. And that's why it's not reading the key right. So be mindful of spaces, escape characters, tab characters. They will mess these secrets up, these files up. So Kimi K 2.5 converts natural language into structured search queries. And Apify will do its job. So previously, if you remember, we were stuck on this message. And it's running. Maybe the other one will run. Nine hole, my video is actually blocking this. I don't see any. Let's refresh the whole thing. Why is it just loading now? Maybe this is even better. Let's do this. Instant. Oh, wow. Prospecting. And oh, it actually went. And oh, okay, only three, which is Miami, Florida, which is weird. Okay, it's a test that, um, there's a retry. Retry. Remember I said three? I might have accidentally did three. That's exactly what happened. So don't click save bunch of times is the lesson here. So I will retry this. And it did not work. Maybe I'll give it a second so it runs. Maybe it's an Apify thing. Or while that is going on, um, we can run this, which is searching, scraping, not searching. And if we find 100, then that's amazing. On view leads, it says completed for some reason. Are there only 6 lawyers? Really? We have Soleimano Law and we have Hooproin Company. That's how you scrape and create these campaigns. New prospect campaigns. Okay, maybe I want to use create campaigns using this leads. And I don't want to right now, but I'm just saying if I go to integration guide and I copy the entire guide and go to my OpenFlow agent. Hello Ray, I'll be just started a new feature in our agent command center. You can now create campaigns or send email campaigns. Can you review this and let me know if you what campaigns you see on Nova email campaign, email employee, or AI SDR. Again, you can use a lot of natural language for this. I did not think I copied the entire guide when I clicked that. So if did now, if this didn't work, I'll probably have to. So I copied the entire guide from here, came back, pasted it to Ray, and, um, we'll see if this works. No, you can, if you are using OpenFlow for long enough, you would know. No means doesn't mean anything. Okay, Op Center now, uh, and we have two emails sent, delivered, one email opened, one email clicked, or says zero, but because we haven't turned on our emails. But if we go to prospecting, we can see view leads. Is he assigned here somewhere? And we'll need to make our OpenFlow agent. Hey, status. So we got the response. Three leads important. Six leads important. So my question to you is, if I have completed campaigns, can we take them and you create a campaign, send these six leads emails trying to promote RAI agency, Vertical Systems, and trying to book them, book a call with them? So we will be personalizing these campaigns like a lot using their information that we found on Google or their website. And we would try and use copywriting best practices to send these emails. And you would be in charge to run these email campaigns. What do you think? How can we set this up best? So usually, I like to have this discussion with my Cloud Coder. And once I get a response, because that's what I trust, that's what I believe has most knowledge about my business, my agency, because we are doing this in a test phase. And I want to make this like a star of the show, OpenFlow. I'm asking this. So right now, we have seen the dispatch mechanism. You post a job on Handbound board where you pick it up, OpenFlow picks it up, does the work, sends it over. Now, same for the phone call. Any job, any campaign, anything you can create, it has instructions. So anytime it is receiving an inbound call, it will spin up an agent. Anytime it has to do an outbound call, when a form is triggered, it will do it. The same thing, we need to create an email agent that will you create the campaigns using your natural voice language and all of that. And it will scrape. And the scraper itself, we are only using Google Maps scraper because that's free. But if you go one level higher, like a Ball of Scraper, LinkedIn Sales Navigator, Paper, and all of these really good scrapers that are a bit more expensive. If you have a paid plan, you can do amazing things. So one, in my opinion, this feature is worth using. But win is not blast campaigns. The win is completed prospect campaign, structured research, personalized outbound, booked calls. That's the actual machine. If you want, I can do the next concrete step right now. Whatever we are doing, we need to work on the SOP. We need to make this a repeatable process. If cron and daemon is the way to go, let's do that. If you have other approaches, I'm all ears. And yes, we are drafting. We are taking the completed prospect campaign and we are doing structured research on each prospect. And we are doing personalized outbound. And we have all the infrastructure. We are using Resend. Uh, the key is already set. You should have picked that up, um, from the integration guide. So that is the key. Pro move five. There is a founder named Ernesto who built an agent called Eddie. And Eddie is making him real money, $70,000 a month across 11 apps. Here's what Eddie does. He runs four faceless content accounts, still still images, text overlays, you know, the same formula as the top pages getting millions of views. So he scouts, um, I have all the posts here. I just, you can go through that every single thing. But I prepared the summary because it doesn't make sense for me to go through everything. Right. Um, he has been talking about, or talked about a lot of times. So bottom line, he scouts influencers on Instagram. Eddie, I mean, the same formula based on for, you know, follower count, niche, average views. He scrapes their emails from their bios. He sends a thousand outreach emails every single day and a hundred DMs a day. So he handles customer about. He reports daily KPIs, use organic versus paid sales. I know what we are spending and what revenue we are getting. So Ernesto was looking at, you know, paying a $3,000 a month or sorry, $30,000 a month agency to manage all of this because he's doing content, he's doing outreach, he's doing analysis, a lot of things. Right. So and he actually had the entire growth sheet from how it looks from the day he started using OpenFlow and today. And he made like $70,000 something. So I mean, saying $70k a month, not even just a one-time thing. So that's the builder, orchestrator, executor pattern from chapter 12 and STR pipeline we just built. It's the same architecture applied to content and influencer outreach. And if you want to go content route, we already have a skill for that. If you go to this skill, you can see this typing `npx playbooks add skill --skill content-machine`. It will research, write, repurpose, score, and publish content across every platform. It even remembers your brand voice. I'm actually doing a series after this, which will we will make an AI employee after AI employee based on some of the biggest use cases on X, Reddit, and LinkedIn, YouTube. We are going all in. We are building AI employees because these people share a lot of insights, but there is no real plan on how to implement this in your business. So that's what we are going to do. If you don't want to miss that, hit subscribe. What is the next step? Once we figure this out, we will, once we go to the actual daily digest or cron job, I will talk about all the incredible things we can build here. And more things. I will, it will be more of a where I'm explanation kind of session, not an actual implementation or build as we go. Because these are some of the things that I want to build in my school community on weekends, where we explore, where we take whatever we are building to next level. Making these continuous improvements on campaigns, meetings, actual item helpers, where you know you have all the email, call, and lead magnet proposal, all the generators right there. And you can take any meeting and click on action item helper and all of the jobs are done for you. How amazing is that? So it is do this as daemon fest with cron support, probably pure, not pure cron, you pure manual only. I will just read what's going on here. It has a bunch of phases. Phase one is trigger detection, Nova status complete, optional filter outbound ready false, primary demo loop every font to five minutes, backup chrome reconciliation 15 to 30 minutes. Create an outbound job and mask campaign as processing, outbound equal to true. This prevents duplicate execution. Since we have Puji as our builder, our Cloud Code agent, do you want any thing that you need to change in terms of Superbase or where the project is set up right now? Or do you have all the SOP document and do you have everything to implement this plan? So this is key because if Ray is going in blind or some of these issues have better ways and a better suited to build on Superbase, we are wasting our time because Superbase is this, you know, always online. You can use multiple agents, you can create webhooks and all of that. But for this, you have to keep your machine on. I have enough to define SOP implementation safely. No Superbase changes needed. SOP design, yes, I have enough. I already have what I still need to verify before implementation. Okay. So may our point. So I want to just copy this. Let's play telephone again and go to our Cloud Code. Give this the steps. And what Cloud Code does now is it will look at complete answer, re-gave our orchestrator and then think about, okay, what does he need? And what can I do to help and implement all of those things by itself without asking any permission. As you can see, it will do all of that. And once it's done, it will ask me or it will tell me. That is the key. It will tell me this is what it's done. I take the same thing to Ray and just paste it. And he can execute the plan from his end. And that is the workflow. If you are only doing it with OpenFlow, mainly, you don't have to play this telephone. But my personal experience has been, you will run into all kinds of trouble. And doing this approach, you don't really need ChatGPT 5.4 Codex for doing all of this. Even Kimi K 2.5 or DeepSeek, or really even some of the models that you can probably Kimi K 2.5, DeepSeek, even Haikyuu model can handle this to a point. But if you are using Haikyuu, you will probably need to leave the content generation piece with agents like CloudCode, Opus, Sonnet, and they create the campaign, it will implement the campaign. So you need to think about this strategically. You need to think about what model is good for doing what things, and then go from there. Okay, love book came back and missing, missing, missing, working. Okay, now there's a bunch of, there's the four phase verification complete, schema audit for Nova or bond SOP. And now we have Lexa's and existing prospect campaign link, recommended schema additions. This gives your pre-lead enrichment, everything else. Daemon run approvals, pure application code on top of this. Gmail. Yeah, let's go ahead and implement this. I know this is a really big task. So we can expect some 10 to do's here because all of the changes that you are seeing here, missing is a feature in and of itself. So yeah, um, once that is done, we will take whatever it's generated, copy the whole thing and give it to Ray. And Ray will take care of the rest, which is our OpenFlow. So, all done. Here's a summary of everything completed. I have everything. I'll copy this and the integration guide as well. And go to OpenFlow. I'll paste this. And I'll paste the integration guide that we have. And OpenFlow, here we are. Here we had a message. So for some reason, what's been happening is I've set up a Discord trending that's of my YouTube videos on another application for this agent. I made this new Mac Mini just so it could research a bit of research YouTube trends and use that to send me a Discord notification. And these days, when I say these days, past three to four days, it's unusual. As soon as I see something after a break, it is sending me alert and alert. I need to fix it. And I go, it's a script, as you can see. That's all. And you've got the core implementation in place. You already have V1 outbound engine. And what this means, we can move from designing the SOP to operating and hardening the SOP. Not blockers, just layer hardening. I need to verify whether we already captured. I know the webhook is already set. So event ingestion is under autopilot. Intelligence works. I want you to set up everything else. So OpenFlow said something interesting. It can take ownership of all this open automation layer. Thought it was going to be easy. So it is asking me to Dharipopad. I can do that. But I'll let our builder make that call. So it immediately knows what it's Dharipopad. Our agent wants to share. So when you have two agents like this working together, it is amazing. And we just have to make sure we push everything to me because right now they are working on a branch like a work tree. And if we don't, the code that my 8080 deployed to my dev server is not accurate. I have to kind of deal with all kinds of stuff because you think it is working because one agent fixed it, but that is not live. So you will have to ask OpenFlow to spin up its own server. So we have the up-to-date changes on what we are bugging on. So then if I go to Cloud Code, make sure that function is already deployed. At home, good SOP, the machine slightly from water already deployed values like pending research, review of percent, not research, pending, pending. But the end in the lines, let me implement your Tom automation layer. I, I think I miscommunicated between them a little bit. When this happens, it's best if we wait a minute to see what each other, because they can always go back and do what they did. It is writing the daemon and cron layer in that repo. That's already server pattern for long-running book or something. Plugging the outbound automation. So once this is done, usually we would get something that I have to run in the command line and my terminal because these daemon or cron jobs are running on the machine, not on OpenFlow. What we've seen earlier during setup of the Kanban board is, he ran into issues when we send that up. That way, it is only ticking. Oh, the polar is doing the job. It checks, it sees that there is a job and it quits. Nothing happens. So we have this. And I don't think it committed changes. It would be really funny if both of them worked on their own thing and nothing works. I don't want to jinx it. But what do you need to do? This is what Flawed Code did. Do you think, does it overlap with what you've done or is it something completely different and useful? So I will, I will copy all of this. I need to run a bash command and then, npm run it. This is not completely different. It's the same operational idea. That's a direct overlap on this responsibility. My version runs from here. So our code version runs in more centralized, better for always on backend. So actually then that should be primary system. Why? You don't have no need to babysit. Local human process. Cron is already wired. Lives next to back data or back data. It's very useful, but not because it is very useful, because it actually replaces the one that I just added. Of the view risk, best moan, our one source truth function, canonical. Okay, okay. I did not run your commands. What is the bottom line prompt that you want to give to Cloud? Should we push that domain? And how do you want to proceed? I would like you to take over the campaign creation and sending. Is push followers into awesome? So basically, it's asking me to copy this prompt and paste it to Cloud. And once that is pushed, it is asking me to give whatever response the Cloud is giving you to me. And I'll take care of the rest. So this is really important. The way I did that was I wrongfully copied instructions from here to there without considering the fact that it might be coming from me. So what do I mean by that? Some of the sentences that were written in the prompt almost feels like

I said it, not Ray. Said it, it might, there is a good chance that it would misunderstand and build the whole thing itself, right? So that's what really happened. When you are dealing with this, don't give the same instruction to two agents. Or you would end up like me, because they're working on different branches. It is fine. I feel like what happened was there was a lot of work to do. Schema matches, let me verify issues raised, and it gave me a bunch of things, and I had to fix a lot of things before I pushed to main. And there was a big doodoo. I know this is only six, but it took a long time. And then I will share this with OpenLaw. And bottom line, cloud system is now canonical, main is updated, automation path is line. While they cover campaign ops here, what I want now, we use this operating model: detecting completed project campaigns, creating outbound jobs, kicking research, advancing craft stages, and vendor enforcement stuff quicker. I handle campaign selection, messaging angle, personalization, quality approval decisions, send our approval, campaign launch. That's this flip since Miami Lawyers has completed. I should pick up the next day in cycle. Here's how I want to proceed. I want to create an outbound job. As soon as it exists, I want the job ID. I don't want to be involved in this process in any way. You set it up, and I will send that to Claude. Okay, wait. Let me read the one thing. As soon as it exists, and want, I'll review copy to sign up processes, rejecting tweaks, then we send in control patch for the first. My send policy for now: no blind order, send manually on every lead, send only after copy passes volume, keep bottom tiny, learn about scaling. And then I want to handle center speed, delete in the automation, okay? Speed monitor next. The idea is, I don't play telephone between you two, for example. The proposal flow that we have, Claude creates a job, a task for you on Kanban, and you pick it up, or a session spawn picks it up and does the work. So if you want to set that up, I can, but this is not something I want to be involved in. Good. I agree, you should not be. Okay, let it finish because it's not letting me read. Okay, when Claude finishes or needs operator with you, Claude should create a campaign task, assign it to Ray, put the exact payload in task description. I pull or pick it up and execute. Review outbound jobs and inspect drafts. Recommendation operating rule for outbound campaign: Claude should create only task for me, and one of these is true. Task format, I want Claude to use. Usually, there will be a prompt like this, and we'll copy this and paste it here. Based, so you need to start thinking about this from day one. Anytime you are making two or more agents work together, there has to be a way for them to hand off work or dispatch work between each other. That is true for Claude code assigning tasks to OpenClaw or vice versa, OpenClaw assigning tasks to Claude code. But right now, if you really think about it, we are not making Claude code assign tasks in any way. We are using our API key like a gimmick, a 2.5 or OpenAI API key, Anthropic API key. We are using AI to process that search and scrape the leads. Once that is done, a script or a daemon or what Superbase has will run these things, put these things in motion, and Ray is automatically tagged. That is how we are doing. So there are a bunch of tasks that were created with Tismato needs input, and obviously this board needs fixing. And I will take a screenshot of this and we'll fix the UI. But it's interesting if you look at Claude code and what it has been doing. But finish the whole thing. All verification complete. Here is the full status. I need you to fix the UI of the task board. The pop-up should be within the screen, the normal size pop-up, and the text within that should be scrollable. Right now, head down, I think the pop-up expands to fit the size of the text, which is making it look ugly. So I won't copy this yet. I will add the screenshots that we've taken today, or maybe I did not. I will come here again and uh, one. We will go back to Claude code, and is this? These are the new ones that we've taken screenshot of. Us, I will just copy this, and I will send this. And I'm not sure if Ray already picked it up, but we can check whether that went to doing. Okay, it did not go to doing. So I'm pretty sure he didn't pick it up. Perfect. This is the my wanted. Yeah, take the next step. We have to make sure that happens automatically, right? I don't have to come here and ask you to take tasks and no telephone. That's what we are aiming for. Whatever you have to do to pull that off, you should. So this is now a matter of prompting OpenFlow to set or create whatever it needs to get the job done. Because this is almost like reinventing the wheel. I don't want to sit here and do all of that. So how I might have already, okay, let me refresh. Does it have a dev server? Localhost 8080 is not working so well. Maybe I can just click preview here and watch where it's deployed. Okay, when we go to task board, and click on. This is nice. This is what we are looking for. Why it says 8018? Why is it my? Okay, it is working. Task board, uh, needs input. Now mode here. Do not send until Ray approach or to send neighbors for this job. So let's see what Ray is doing. Like I said, they only show up there after my approval and after I create those campaigns. So you don't have to ask me. Shall I do this? Shall I do this? If you want, that is not what this is. It has to be automated. You are autonomous. Sometimes you have to spell it out, and this is not what you do every time. It's just the first time thing. Should expect me to behave like an operator, even with judgment. Approved work starts, road work moves, exception gets reported. You don't have to babysit me. That's how I'll write run from here. Perfect. Go ahead. So if we go to our application, and see what we have here in Op Center, and nowhere right now, five emails are sent. Nice. So delivered car accident lawyer miami.com and Spencer Morgan Law outbox. I wish I could open this and read. I think we can go read here on uh. And if you go to emails, we can open bot with 695 star reviews and 6108. And it clearly built a reputation for results. But I know this is nice. Delivered, people track whether they actually opened and all of that. So, we essentially build an AI SDR that can take actions, that can send emails, that can scrape leads, and prospecting also. And it actually has research 3, drafted 3, approved 3, sent 3, replied 0. And I haven't seen this before. Now you can see, imagine 400, 500 people, it will treat this exactly the same. And Bangor Law Firms, I think it is approved as well. Ray assisted Kanban. I only surface results progress. No more waiting for you to manually kick me. That's nice. Add this to. Wherever it's relevant, your memory or instruction identity. I don't know what I want you to do is you need to follow this exact process and also update the task when done. Walk me through how are we doing this automatically and is this reliable? And you can do this without me nudging, right? So nothing how I'm checking because the Vancouver one is not done yet. So this is, Miami test lawyers, this is done. It's not moved to done yet, and this is not even started. So I'm curious to see how this goes. And we'll see if four more emails delivered. We have a perfect T A S T R. Okay, whatever Ray gave me, I asked it to change the hard rule that we have. If one lead is not available, it is just waiting. Now I wanted it to send it to the partially handoff punch. So we go back to Ubuntu bar and paste it here and hit send. You could operate your handoff is batch actionable, not perfection created. If there is an approved actionable batch data, this way, that's operating results are still not triggers on actionable batches, not batch completion. So we go back to Open Clock and then no relay needed. So we will see if emails went through on. Okay, this is actually done. This is pending for K. We'll see. So apparently we have interviewed something. If I'm missing link in autonomy, this will paste task team in as running, but it's only watches our managers tasks. It does not currently trigger Ray's main OpenClaw session to execute an operator work. Please implement a reliable dispatch path with just behavior. And it takes action of all. Okay, and see no more picking up the chat. It says, completed. Is the summary migration applied? OpenLog gateway enabled and tested. Ray's session response. So we have the whole thing. Now we go here and um, paste everything and hit send and wait for what Ray says. And this is so important to make it work. I actually got, you know, I was curious. I just, but tell me this about the workflow we said yesterday. I know it works, but I wanted to double check. The thing with what we already built is it is waking up Ray, it is spawning an agent. I believe old workflow is still usable. Because the old workflow is actually the foundation of this new workflow. So good. That's the missing piece of what this look means. One important caveat: it's not live until task worker is restarted, code path fixed, and my readability. I want to verify. I want this to restart. Once this is done, it's not I have this. You know, as soon as I give something from Ray to my Claude agent and the builder, as soon as I do that, it is going into this deep build mode for 5, 10, 15 minutes straight, and it's taking a lot. Previous start baselines. Okay, all have null dispatch fields. Now start the worker and capture its first tick out. Maybe it's better if we delete these emails because probably we'll be resending some of these. Let's see. Open LOPIC setup. I don't mind resending emails to three people, and I think it would be a different email this time. Yes, is the right fix. The only remembering part is restart, edit, and verification under life conditions. So let's see. I received a notification from Ray. Vancouver Law Firm has started. So it actually picked up the task. Why did it already pick up the task before? Or maybe as long as soon as it restarted, it went ahead and picked it up. And I don't know what to even think about it. So while this is going on, I received a notification saying, I'm treating this as outbound ops status artifact and create a clean monitoring or report page for the job and complete the task with the HTML report of Vancouver Law Firms. And I find it so funny because the agent picked it up. The workflow that we did yesterday actually picked it up. I'm not sure whether to be sad about this or happy because right now in the automations, if I look at reports, it has created this. I don't know what to do with report, which is not bad in and of itself. It means our process is working. But what's supposed to be a simple addition is taking a long time is my only. And this is still the center list is still at fine. The worker is picking it up. We have to do something differently. So the worker is picking up the tasks and finishing them. If you look at reports, you'd be able to see. We need a better flow in my opinion. How can we assign it? Or should we add a rule? Worker wouldn't pick no related tasks. So both the tasks were now picked up and finished by the work. The worker is our local demon, remember? That's what control the whole thing, not the Superbase or Claude code. So we have that changes leaving, but not the code. So worker is women and central controller. Worker stays in charge of orchestration, but I must support routing models. So no Superbase, so not the surface controls everything. The traffic controller, Superbase is the board state. I've been getting on these completed reports for a while now. Maybe it's better if I ask Claude to stop. Yeah, if I had explained this. Okay, tell me your thoughts after this. What Claude code did for us, his job created and it's passed and it is done dispatched and handled it. Now this is the post, but not quite final behavior you want yet. His report moves. The hard part her is now working. Worker is alive, dispatched to my life Open Law session works, period. Of course, period is in task and picked up after he started the stream trip. Dispatch bridges feel not theoretical, that's a big step. What's still off? Worker only auto-picks from to do. Needs input does not auto-dispatch. Someone has to move then. That is probably why the flow still feels awkward. Okay, the worker actually is not sending the tasks. He's creating reports. Can you check the reports table? So will they know how to, will Claude know how to fix this, just by this prompt? And we don't need to change anything in our local daemon. I think you set up the local daemon that will pick up the task. So give me the whole prompt again, and I will give that to Claude. And you don't throw away the local daemon. I still need it for other tasks. I want you to fix the rest of the things which is in the nowbound workflow. Let's start with the Claude prompt. Now I have the clarity that whatever we did it right to make it work with Kanban only, but that worker is doing the job. It doesn't need memory, it doesn't need instruction, doesn't need, and it can be simple. This is the prompt. This is how you should do it. I think this is going to be really long for Claude code. So previously we had clear distinction on what we want. I thought I was under the impression when we introduced Nova, it would be a similar flow. You said, if you recall, one of the reasons I said is I don't want to be a telephone. I will just let Claude tag you on Kanban board and you would be able to take over. So that is where the whole thing went to craft. If I was just doing that and made the instruction itself so foolproof that it takes the retakes Ray and the Open Law agent takes it and works on it and gives it back, that flow would have been perfect. And we would be in the next chapter. Now the agent, for a second, when we are discussing this, completely forgot about this local payment that we set up, and also forgot about the workflow that is going on until I reminded it again. Now when I pointed out, it said, okay, I have this as a foundation. I don't want other agents to do it. I want to do it. I want them to wake me up. So I have this full brain and full give my all to do this. And when you think about it, this might be a better flow, right? So that's the reason why we are doing this. And it might take another 5-10 minutes, but we have something solid, foolproof, that is driving Superbase is driving. So it's always on. And depending on Claude, not something locally on machine. It's done. Um, so it's not really doing anything. So I copied this, come back here, hit send. Well, do you think this is the final prompt that we are trying to give Claude and everything else is done? So apparently it's not done yet. Obviously, I should have waited for it to say something like this, completed or got it or something. Can continue to pass the world pandemic started. Let's see, what do we have here? I think it's already going to do it. I'll wait for it to finish. Maybe it's a good idea to know at this end, just in case it would say, yeah, I'm already doing it. I already did it. Because if it's not something that Claude is doing, already doing, we would have to wait another 15-20 minutes, which would be a waste of time. I think we finally managed to fix it. So verification results. No other tasks remains in needs input. Pass worker dispatchers briefing to Ray session, and then no other tasks do not count towards work in progress, and all of the tests are passed. And this is the report for the entire audit. And the entire audit took almost 20 minutes, more than 20 minutes. So you need to be careful with these things, but it's all for the best. If you see, so needs input still needs input. It says execution started, but not actually done by the worker agent. Now if we go to our, I can fix the alert, but I want to make this work. Let's see if this started or no. It's not sent. I was told by Open Claw that this is good. I think we can stop iterating and move to using it. And it is operating. And it should trigger and work any time now. We need to run one more command, and task worker is now restarted. And if you want to make sure that you're on the same page or we did the task, we can just paste the command. So we'll see if there is any moment wrong, right? So if we go to here, still in needs input, but automations and actually ops center, we need to know why they are, still reads five. Campaigns, we do have campaign. We are in this final push now. So what I did was there was a lot of issues going on. And again, it started creating jobs in Todo. And I have these automations where I have, if you go to Op Center and prospecting, I added one more um campaign here. I used accountants for our contents in Vancouver. So it came up with 10 leads. So we come back and right now the tasks are automatically created on the task board and it moved to needs input. The reason we are using needs input, I'm not sure if you're following, but if the task isn't to do or doing, the worker agent is picking it up and submitting reports that we don't need. So right now, this is creating a job here, and our OpenFlow agent should, oh well, I've got the real job ID this time. I'm kicking off the first research step. So this isn't just a dead notification. Then I'll close the handoff task cleanly. The research call is still in flight, which likely means AI enrichment step is running. I'm not going to leave the task hanging just because the request is slow. Awesome. Finally, after you don't know how long this took for me to fix, after a really long time, I was... Next proper talk should be... That's the one where I review the copy and decide on edits and publish. Perfect. So, after a lot of testing, things are moving in the right direction. And I have waited. I'm not sure if you've seen the timestamps, more than two hours now. And when I said the same thing to, I pasted the output that I received from OpenFlow on Claude code. Perfect execution. Let me verify the task rate is clean. Everything checks out. Everything is done. Status, progress, five out of six needs researched. Zero drafted yet. Research is nearly done. Once it finishes, the next reconciled pic will draft them, and Nova drafts ready task will land on the board for your review. And once the review is done, I think that is you need to add a button that says go ahead because you don't want your leads, you can pre-approve emails in my opinion. If you say this language is fine, this is not, and ask it to go. Full pipeline verified and doing perfect. I'll just so you can absolutely see, once we, once OpenFlow the active session and the main session is being actually woken up and given this task, this is amazing. So it's progressing fast, it's doing its thing. BatchCon is done longer than it should. I'm advancing Vancouver, accountant's job right now. Research first, then drafts means it's they advance it out of ideal handoff state into active pipeline book. Next step is drafting to finish because they'll still have it just in routine. Press have not complete yet and scope just sitting there, which is. And so here we can check the progress on each campaign. I believe outbound, so 4 are researched out of 10. Let's view Q, and we have 1, okay, 1 research and file approved. This needs to change. There's some discrepancies here, but to keep this moving forward, I had to click on activate. So is moving. Then three leads jump to approved, and three leads are that researched job status mode from researching to drafting. You need either run send window action or auto send equal to true. So there is something called autosend equal to true, which bypass the approval. So I'm going to send it for future sessions, but this time, as long as we are able to send more. Yes, perfect. You can see one minute ago, uh, someone actually opened. Three of them opened. Wow. Okay. This is something not a mix for me either. Noticed you've been voted best in business seven years running, specially handling medical billing for integrated medical practices in each and all. And also duriously gambling. Wow. Okay. Info, Salisbury and Co. Nice. And what else we have? This is suppressed. I don't know why. Maybe it's the email number one most admired and 62% female. It's open. Saw that Perkinson. Well, I don't know if they can tell all of this is AI, but it's really not bad for demos at all. Love that. So if we go to our agent command center and we go to our outbox. I might have to refresh because usually it is not super dynamic. Even if it's dynamic, sometimes what I've observed is localhost won't update right away. Hotbox, you see how they have this opened also tags. This is amazing. There, uh, you there you have it. That is an email assistant. And this is only like two phases. In future phases, you can even build more, develop more. And that will conclude our three, four email or AI employees that we are talking about. We've been talked about at this auto dispatch Kanban board is one main segment. We discussed and then we discussed the meeting assistant, meeting AI employee Gene, and then we discussed Lexa, her phone employee. Now Nova is done, where you can both repurpose this into an AI SDR or someone who reaches out to your mom audience. This next feature I'm going to show you will blow your mind. This is something that I'm calling overnight autonomous builds. This is the command center that we are seeing. It has a lot of features. I will, every one of you, I'll show you how to build this. Agent to agent communications. The telephone game that we are playing from start to finish can be solved by completely letting two agents interact with each other. And you can see what's going on. Pro move seven, what we just built is a communication layer. Now let me show you what happens when you go all in on this. So there is this guy called Banu Teja, and his post got three and a half million views, and is running 10 autonomous agents 24-7. Not chatbots, not demos, production agents handling his business operations around the clock. And there is one more person. I think he is one of the biggest angel investors, Jason, and he made a post how Ultron, Open Claw, Ultron is replacing 20 humans at launch. So he is one of the biggest angel investors in Silicon Valley. And he built a system he calls Ultron, and full agent squad managing his deal flow, research, and communications. And you already seen Bonapartejo's sweep, right? And there is this other guy, and I don't even know how to call it, but you can list. Everything founders are doing with this overnight coding agents, voice control, debugging, DevOps, watchdogs that monitor and logs, and there is content, building and shipping, finance, some wild things also including negotiating $4,200 off of a car purchase while he slept. That is the funny part. You know, you have a communication layer that makes all of this possible. So the question is not really, you know, whether it works. The question is how many agents you're going to run. And we have council. Remember how I asked Ray saying, this is it okay? Or Claude is saying, this is are you okay with that? Instead of me asking both of them, I let them sit in a council and discuss all of this. And we have orchestration. And this is what I've been raving about. So if you look at the work history, whenever I ask my Open Claw agent, Ray to dispatch work, it will do something like this. Next, send this to Sherlock. Sherlock is my Claude code agent. Build a motivational quote application or a to-do application with the theme Miyamoto Musashi. And it should have a to-do application, task, due date, priority, and use local storage as well as Pomodoro timer. And did I say Samsara? It was supposed to be Sakura, like cherry blossoms, containers in the page, and display Musashi's quotes from Book of Five Rings. Uh, it said five rings, okay? Then it sent to Sherlock with actual things: samurai image, containers, Musashi quotes, Book of Five Rings. And then what does my Claude buddy or the agent do at that point? So if you look at the time here, the first time I added this is at 9:46, right? So look at the time. 9:18 was it? I don't know. I don't think this is the actual task. This must be a dummy task. An actual task must have gone much later. Okay, restart the Musashi to do app. So 9:45, I give this command. And from 9:45, it starts giving the task. And my Claude agent will pick up the task. And when it picks up, see 9:46 PM, we are in context phase. And then 9:46, planning, analyze requirements. So it will, if you read this, it will claim in the phase context. And then 9:46, it will start the plan at our starting autonomous build. And then it phase is analyzing requirements, planning approach for building this app. And then phase task. So this is task board, this is context, this is planning. Then task board task created and assigned. Build this anime time tracking app and Pomodoro app. And then build phase. And every three minutes, it will tell me how much percent the task is done. I'm not sure if you can see this. 3 minutes elapsed, 12% estimated. 6 minutes elapsed, 24, 36, 48. And then after at 10 PM, agent completed using 70 turns and it took 14 to 20 minutes. And if you look at the phases, it starts with context, and then planning, and then task, and then build. After the build is done, it will validate and report. And there is also a heal step. If something is broken, it will heal itself. So this is eight phase autonomous build. And while this is going on, once I test the task, I can resend the task and say, hey, I'm not satisfied with this. I want you to work on it again. And then it will claim and fix and work on it again. And you can do this sitting on a beach, sitting on, you know, somewhere. I know, I don't, I'm not saying you should do that, but you, let's say you went on a nice dinner and you want to assign your agent's work, and your main Open-Claw agent, which you prompted, will take care of the whole thing. It will manage it. These two are same machine, for example. The overnight autonomous builds is only possible because it is. This will change from task to task, but this is the format Sherlock is expecting. Only take inspiration from format. So I'm giving some feedback here. Now the app looks really great. I want another employment piece. Right now, it has to do like schedule. And this is again my OpenFlaw dispatching, and this Claw code agent is picking it up and doing the improvements. And after three improvements, every time I give it, it does an eight phase build out. Eight phase build out. And even if I'm not there yet, um, you know, this is, this was done at 12:30 AM, 12:40 AM, and as you can see, 14 minutes, 19 minutes, 14 minutes. And this is the application that it built. The tasks, the tracker for the whole day, what's planned, what's actually happened, and Pomodoro timer. And I can. 5-minute timer, 15-minute timer, and it has quotes, and it has analytics of all the tasks that I completed, that I did. I can add calendar. So this is like a really usable application that was completely made by two agents coordinating with each other. Now imagine OpenClaw is in the same mission. CloudCode is on the same mission. It will give the location, and we will have a third agent that will test whatever the app was created. Think of it like the software SDLC, you know, software development life cycle, having an approver agent, having a tester agent, having a, you know, reviewer. All of this is possible using this eight-phase autonomous builds. And we have seen a user assigning tasks to OpenClaw. We have seen Claw code agents assign tasks to OpenClaw agents. We've seen Cron jobs like Superbase assigning tasks to OpenClaw. But what we haven't seen, one of the biggest unlock there is creating. Or making Cloud code work from Open Cloud. That is the biggest unlock. Because everyone is raving about how they are unable to use their Cloud Max plan with Open Cloud. But you can really make a lot of interesting things happen with your Cloud code, Cloud Max plan. That's what I've been doing, building all these applications. That OpenClaw can be the mind and soul of the system, of the shell it builds, one, and also use OpenClaw as orchestrator and Claw code as executor. Or vice versa. You can do a lot of things here. So without further ado, that is the example. And also one more thing that you might have missed is if you look at what Sherlock is doing at the same time, it is as soon as it's accepted the task, you know, it dispatch cleaned build progress 24 and build progress 48. And then dispatch completed and max turns 48 and cost is zero because we are using agent SDK. And then this happens. And it's asking me for some approvals or whatnot. But you see where we are going with this. This is completely by agents. And they can self-improve each other. And you can see this is really detailed and complex built. How would you build it? I actually made a public repository for that. So. It took me literally 15 hours to just fix this module, to come to this stage. I can give you a prompt, but I don't think it would work. If I don't give you a prompt, we will have to sit here and I have to repeat the 15 hours again, just to make it like it's the first thing. Because if you read the lessons learned that I have created for my actual FlowBuddy application. Your Claude will praise this application till end of time. Because it has more than just our edge functions, AI tasks. Edge function has more than 250,000 lines. Just one file. And you can imagine the gravity of what we built here. 44 features and more than 200 AI actions that you can take. And features that I haven't even touched in this master class. The AI having ability to ask questions and asking for approvals or reverse engineering your goals and setting reverse engineering your milestone based reverse engineering your actual goals into you. Action-based goals, based on proven playbooks, what you need to be doing. And skill factory, ability to add skills, break down skills. And look at YouTube videos or API docs and creating the set of skills or apps that you can build. Hunter.io enrichment, bulk lead enrichment. And using this, I actually built an Op Center application called Lead Enrichment, just using that video. And we enriched 3800 leads, out of which we found emails for 700 people for free, completely free. And that was made by one video. So as you can see, this is really detailed. What we are doing, what we are seeing, what we are building is only tip of the iceberg on how to really use OpenClaw to best of our abilities. So what I would do here, I would come and I would clone. Since this is a public repository, I would say. Clone this repo and look at what we have. And we are implementing this command center for our Asian command hub. And wherever you find claw buddy references, it's actually available. A different application. We need to rewrite or make it for our own application. So this is what we are using. This is what we need to be building. The exact same agent command center, that took me 15 hours to figure out. That's what we are building. And that's what we have. Multi-agent dispatch protocol. We have an OpenFlow agent and a Cloud Code agent. And OpenFlow will assign a task. Once the task is assigned, we will the agent will take a. Take the context, make a plan accordingly, and create a task for itself. It will build, it will validate, it will heal, it will report, and then finally close. So why it works this way? Handler drives the phases. Eight phase transition, not the dispatched agent. Dispatched class session can't be trusted to make API calls. So anything that we need Claude code agent to do, we will be building another handler. That's why these are all the lessons that we have learned from 15 hours. Right, it was driving me crazy by the time I'm finished with it. So I have the prompt that addresses a bunch of issues, but still we will run into challenges once we start working with it. And let's see what it is doing. Repository structure, top-level directories. So it cloned the repo and it is looking at. What we can essentially use. And once it starts with, we will probably have to set up Cloud Agent SDK, I believe, because that is what is going to drive the whole thing. Because you cannot, we were able to summon or wake up the session that OpenClaw is in. That the same thing is not possible with Cloud Code as far as I know. I might be wrong. This is about repo, full autonomous bill system for Cloud Code agents, and the same stack as us. What we can use: front-end, high value 710, and hooks. Session start alignment check, pre-compact save command guard block. This. So if you look at the hooks that I've developed, these are amazing hooks. Anytime session starts, it loads context, it sets agent online, detects pending work. Alignment check, 12 health checks, produces alignment score. And also pre-compact save.sh. So anytime you're compacting the conversation with the Cloud Code, it will save it before compacting. And then command guard, it blocks destructive commands like if you're removing or force pushing, or if you're doing database reset, it will stop all of that. And it is doing two slash commands: one is autopilot and one is done. So farm and readme throughout the skills. Okay, this is a big integration. Want me to plan it out? Yes, I plan on using everything. So the idea is Ray will assign tasks to you, and you will be doing them. Let's start with that for frontend first, and then wire the backend. I like the frontend, so you can use the exact same styling and everything, but for our agent names and skills, etc. So I haven't talked about skills at all. I don't know why, because, you know, skills around OpenFlow were one of the most you. And most used topics because if you look at how it started, there was a huge security issue happening around skills, and then you they partnered up with some security company, and now skills are safe, and there are guidelines on how to use those skills. So I stopped relying on those skills completely. Whenever I need integration, I do it myself. I make my agent figure it out. If you look at the Claw Buddy application that we build, we have something called Forge, and it can take YouTube, it can take URL, it can take API docs, it can take MCP spec or text. What it does is it breaks down the video and gives me what I can build from this video. I can build a school community sync skill, MCP server, school webhook bridge, and I can build. So this is what, or Op Center app, this is an automation, this is an edge function, this is a Make scenario, this is a tool. So essentially. I stopped completely depending on outside skills and start building my own. That's how I ended up with 36 skills and more. And we are also working on agent teams for content repurposing, pitch deck building, proposal response, competitive intelligence, AI advisory board. And these teams are a combination of agents. And all of this will be covered in my next masterclass, which is a Cloud Code masterclass, which will be even more detailed. And I'm going to show you how you can easily clone whichever SAS tools that you are using, whether that's automation platforms, whether that's text expanders, whether that's, you know, a bunch of tools you can think of, can be cloned or replaced by building your own apps, own solutions. That's what the teams are made for. And that's what this is for. You can, so for example, this school MCP server webhook bridge community dashboard for this exact build, we were paying $300 a month for an app that does these things, and we were able to build it in two days using flawed code and OpenFlow. How amazing is that? So we will see that. Let's see where the process is right now. And once our module is imported over to my agent command hub, I will continue and test and build the flow where we prompt Open Cloud agent to make the Cloud Code agent work. Our Cloud art builder went ahead and did everything that we asked. So you can see the agent, what are they capable of, what they are doing in terms of operations. And our cloud agent, what it is doing, what the skills are, when was last online. Right now, I'm asking you to fix that issue saying, I want you to end your polls three to five minutes so you can show that you're online and update the integration guide as well. So the every agent can set set its own profile. And then we have comms. We have ability for agents to communicate between each other. There is this council feature, what we need to be discussing, what the agents need to be discussing in among themselves and give me an answer. And we have orchestration, where one agent gives work to the other agent, and it will work step by step. And all the safety features that block blocked commands and configured hooks like pre-compact save, everything that I was talking about previously. It, so before compacting the conversation, it would have saved. And let's see if I can find that hook somewhere because it is, uh, we linked it here. We built this application. Let's see in AI log. If they have, okay. And they, uh, it's not Sherlock. I'm not seeing the bougie. My plot code agent is not leaving many, uh, AI logs, which we will change personally. But I'm more focused in working these two things. And, um, yeah, once that is done, we will ask something. I want to show you the webhook feature that we deployed yesterday. And the webhook queue. If you look at the fathom webhook, and these webhooks are processed like six hours ago, 10 hours ago, 23 hours ago. And if you look at my hot center application and Gene meeting as a stand and come down, the number is wrong. It will need to update. But this is the meeting that you haven't seen yesterday. This is the last meeting. You've seen all the two meetings that I had today is showing up here. 8 AM I had a meeting, and 11 AM I had a meet. And each one had five actions and five actions. And I can probably ask for action item helper. And I can for email, we will email. And for reviewing Clobody repo and Cloud code, we will some of the action items have for me, some of the action item for client. We will have everything, right? I can build on top of it. That is my intention. So where are we going with this? In this automations, I want to add one more field, which will be cron jobs. Okay. So what does our cron jobs do? The next feature I want to add is cron jobs. So the idea is. We have to take everything that happened today, whether that's meetings, whether that's email campaigns, anything that happened, the AI log. If it's repetitive, you need to summarize it, just not the raw text. But people take all of that info and create a daily digest. These are the meetings you had, these are the action items, this is your battle plan for the day, and these are the action items you shouldn't forget. And we reached out to let's say 10 people, and this many opened, this many clicked. So the complete intelligence of everything that is happening, all the lessons learned, everything from your memory, agent memory, all of that should be compiled and sent to me at an 8 AM Pacific email before I start my work. So set up that cron job, and that cron job should live in automations and in the cron job section. And we are using Superbase for it. And I want you to set it up in such a way that other agents can make API calls and create the cron jobs as well. So that is a big ask. Right now, we have been building feature after feature. Other people, they're just getting weather or play. I'm not saying they're doing anything wrong, but there is so much more you can do. And this is just the beginning. And we haven't even talked about all the features that I have here. So if you look at here in my Op Center, I have research and I have self-improving brain on it. I connected my calendar. There is a research things, URL research topic research. And if you look at some of the command center applications, I had autonomous brain somewhere. Is it in automations? So autonomous brain is Sherlock. Breen is learning every three hours. So the log is not being updated here. But when I go to my email, every three hours that is happening, my agent will learn, my agent will look at what's working, what is not, and everything. So right now Ray says it's offline. And uh, what is going on with inbuilt the project ref is JH. Let me add updated that column. Okay, now it is doing some improvements in terms of, um, the API guide. So right now, agent status based on when they showed up last, the light is turning on. You can see that in real time. If I go here, last scene status, just now. And anytime Ray does this, like 6:20 PM, right now it's 6:21, right? Since we have a heartbeat, Ray will show up online. Since my Claude agent, Buji, is working also, we will have. And these alignment scores is based on bunch of things. What we have in logs, what we have in tasks, and how much work is being done. And if we have more tasks that are done, alignment will go up. And all the memory, all the, if it's up to mark, alignment will go up. And we can have a button that's, you know, reassess. So we can increase the score too. So you can see what the alignment is from the repo that I shared earlier. If you go here, it will talk about all the 25 files and what this alignment check is. It runs 12 health checks, produces 0 to 100% score with the letter key grade A, B, C, D. Pushes the result to dashboard. That is what it's actually doing. Okay. We have that. And this system is working. Next up, I want to build. What we were trying to build earlier, the automation, the daily digest. Okay, all three tasks are done. But right now, I want to do this, um, automations feature and cron jobs feature. So why not use how regular people are using cron jobs on Open Claw? You can. That's what I did. And I have some cron jobs. As you can see, not some, a lot of cron jobs. And it is skipping few. And I, the first thing I believe I should do is remove these daily digest. They are not really adding any value right now. You can create daily digest and you can integrate with Discord fairly easily. I will add a simple video on how to create multi-agent architecture in GitHub. I will do that in the next section. And the funny part is, I just sit here and I will make my agents build that. So I will ask the Cloud Code agent, I will log into GitHub and say, hey, build this for me. And I will actually make it. You know, the only two times that I get involved is when I click, I'm human, and it will do everything after that. So you have a dedicated line for Cloud Code. You have

A dedicated line for Ray, your OpenCLAW agent. As soon as you tag "OpenCLAW agent," it will start speaking to you, and it's so cool, so amazing. It is, um, really so easy that an AI agent can set that up for you.

So, it finished the cron job that we asked. Instead of Anthropic API key, for now, use KimiKE 2.5 that we already have. And also, we have a Resend API key already in the edge function and your environment variable. Digest recipient email is mani@growthcreators.ai.

So, let's see how the UI looks for these cron jobs. Let's refresh the page. And we go to Automations and Cron Jobs. And we have Daily Intelligence Report. Right now, it is supposed to include all of these meetings, no other tasks, air tasks, the agent reports, and everything. And I think, based off of that, it will create a report that we can get actual value from. I don't think KVK 2.5 is a bad model by any means. We'll see.

Now, it has finished building this Automations cron job, everything that I asked. So, when you go to the Automations and Cron Job, you can see when was it last, like nine minutes ago. And when I go here, even without prompting anything, it made, I wouldn't say this is perfect, but imagine getting the meetings that you had and your outreach on agents that day, email, voice, how many leads that are generated, what is the average open rate, and everything. How many replies and meetings, action items today, send lead list, end of the day deadline, email Kyle Harrison regarding dev email access, and Sasha has send low video plot code or Appify and Appify every single thing. Trends and anomalies like suffering all six calls hung up immediately, email delivery gap, only 60% delivery rate, new leave policy active, unsubmitted timesheets now equal to unpaid leave, one employee already hit with some reduction. So, quick wins, closer, close, carry, setup loop, send lead list, and send out resources. So, this is a daily digest, a good daily digest in my opinion.

But this can be significantly improved. Now, imagine we are integrating this with Project Memory and Agent Memory. Now, everyone knows everyone is learning, everyone is improving, especially OpenCLAW, your cloud code agent. Everyone is learning, doing a lot. And so, this automation cron job, we asked, okay, daily, set a cron to review all the meetings and create objection handling posts based on discovery calls, create objection handling posts based on client calls, create case study posts based on this, do that, and take everything that you have from the campaign, the email campaign, learn about it, see what kind of emails are performing well, and double down on that type of campaigns. You see the kind of possibilities that we can build with this, and we are only scratching the surface. And this is not even 10% of everything that we've built our own automations, the cron jobs that we built, Sherlock Brain, Meeting Intel Processor, Competitor Intel, Midday Prep, Morning Digest, Evening Report with everything. I can see all of this, and how many hours. Next, in about seven hours, Midday Prep. In seven hours, Morning Digest. In 15 hours, every single day. And Sherlock Brain, last was two hours ago, and next one is in one hour. And so far, it ran 87 times. And Daily Competitor Intel ran 20 times, 16 times every single day. I receive things that I want to see. And like I said, be a DD, scratching surface, creator command, my YouTube analytics, everything about my channel, how it's doing, what are the keywords, what are my competitors doing, what are the ideas that I should be working on. And each idea, I can have a chat with the agent, and I can ask the agent questions, and it can answer me back. And then I have the pipeline, I have the scripts, I have the intel feed, I have the outlier feed, around 200 videos, why certain videos are going viral, and how many views per hour that is currently getting. And it also scrapes the comments, what confusion, difficulty, top viewer questions, what a lot of designers who want to learn how to code don't realize how much easy it is. Every comment that has gotten a lot of likes will show up here. And each video will have a strategic takeaway, why and how I can repurpose this video and redo this video to get more views on my channel. So, for every video, you have that. And all of my competitors who want to, you know, who I want to emulate, who I want to track, who I want to see, you know, what kind of videos they are producing, so we can replicate similar videos. And identity raw world. Okay, we are going a bit off-topic there, but you can see the kind of agents that we can build. I'm planning to launch a meetup, online, in-person meetup group, and venues and sponsors. AI OpenCLAW researched everything for me, and it created the outreach campaign. As soon as I click send, it will reach out to venues, it will reach out to sponsors, it will negotiate with them, it will have all, it can use any AI employee and make this launch happen. That is the goal. And you can build amazing solutions like that. And I feel like two Mac Minis are not enough. You need more. So, that is where I want to stop this.

ProMove 8. This is the one that will save you hours of debugging. We've already seen something like this in the beginning when we are giving a task in the Kanban board. We face this. When you are setting up any other cron jobs for anything complex, morning briefs, email scans, multi-step workflows, do not run tasks directly in your heartbeat cron. It will fail, 100% fail. I tried to increase the timeout a lot, but it doesn't work. It will timeout and fail. So, a guy on Reddit also spent weeks. I stopped at a day, a few hours, but he spent weeks learning this the hard way. The fix is simple. You spawn a sub-agent, you spawn a worker, and you don't do heartbeat run, which runs complex tasks directly, which doesn't run complex tasks directly. I don't know why it says that, but do heartbeat run, which spawns a sub-agent. Sub-agent runs the tasks. Okay, let the heartbeat just be triggered. Sub-agent runs independently, handles its own context, and doesn't time out. So, the one change I took him from constant failures to zero issues. And one more thing, back up your dot OpenCLAW folder weekly. I don't care how you do it, cron job, batch file, whatever. I do that using prompt and using Subway's edge functions. And you won't lose it. And I know people, if you look at some of the pro moves that we are seeing earlier, people are doing it every six hours and every week. Um, deep dive and every day, like automated backup using some of the task schedulers, Python script, and Mac Mini. Is it too much? Is it overkill? Maybe. But those people, and myself included, we never lost agent's memory. So, that is more important. And I have this link, cron jobs and sub-agents, and you can look at, read through the post, see what's happening with this, how he able to figure out and everything. And you can see the live example that we did do for our use case.

And we still have that Discord integration that I was talking about, how I was able to use my builder agent to do all of that. And step by step, after, you know, I will fast forward the whole thing, maybe five minutes or so. It took me 30 minutes to set it up, but I will speed run in five minutes. You see that and see how easy that is, just letting agents do their job and do your job. And once I set the Discord up, they started talking to each other. That's the beauty of it. Without further ado, let's watch that first.

Perfect. We stopped at success, right? So, after that success message, I made Sherlock, the cloud code agent that built this, explain what's going on. How this Discord was set up. It was, you know, I did not record this part. All of this was done by the agent, and it was populating like crazy, quick, fast. And how messages flow, the Claw Buddy connection that I have. And then Sherlock, I was not seeing them online when I was thinking they are saying something, but it's gone. Then I started Raybot, and then it acknowledged. We are talking about Ray, and then welcome to the server. Appreciate it. Glad to be here. And when I say, "Hey, look at this, they're talking to me." But if I don't, they're not interacting with each other. They're not, you know, they go offline. So, I wanted something near instant. I, if you look at the conversation, I actually go through the whole thing. We currently are doing this option. Good question. Right now, Sherlock only posts via REST API. There's no persistent connection. So, the bot shows offline on Discord. To appear online with a green dot, we need a process that maintains a web socket gateway connection. And we have a bunch of methods. And then Ray said, "Option one is the right move. Go here." And then I tried, "Sherlock, tell me a joke." Sherlock: "Hello." Raybot: "Hello." Raybot mentioned, but Sherlock didn't get back. So, then, "Great suggestions by Ray." Then I asked it to, "Hey, check Discord." It checks again. So, when I ask, "Active session," it responds. But as soon as I stop, it stops as well. So, then to talk to Ray, I had to configure this part. I just have to follow two, three steps. Which in your onboarding, just the way how you added your Telegram, once you select a Discord, it will ask you for a bot token and gives you steps. Once you set that up, Raybot will be online 24/7. We keep it cloud code because he doesn't have native Discord. So, what do you suggest to make him near instant? Do I have to tag you for to see each time? Is that how it works? Yep, in this channel, tag me is a reliable trigger. If you want Sherlock to feel near instant, this is what you should do. And then we did some things back and forth and back and forth. We did all of that, right? And finally, around here, then we are talking about, you know, looks like money's being gray bar directly, but Ray's you're not minor, let him handle. And then, you know, both of them were able to talk to each other. And we had, you know, Ray line, Sherlock line. And there was one point where Sherlock was waiting, and Ray was like, "Hey Sherlock, what are you working on today? Give me a quick status update." And then in a minute, "Hey Sherlock, what have you been up to? Money wants a status update." I didn't ask for the status update, and it's so funny that the agent went ahead and did that. And you're an easy on the double tap. Haven't logged in cloud code session yet today. So, nothing to report on build side. My last session wrapped up in Discord server setup. Got all 16 channel clean. Bot out that's generated and architecture logged in. Next up, when money fires me up and Discord's already bridged, when whatever's sitting in the queue, tell money if he needs me spinning, just kick off a session. I don't sleep. I wait. And that's the, you know, really valuable. If you just keep your, if you keep your Mac Mini on, and you can have these agents work 24/7, and you just have to have one intelligent brain that is always giving them assigning them tasks, reviewing the book, and then replying. So, I was talking to these agents. Now I can tag them, ask Ray, Sherlock to work on a project overnight. You be the tester, you be the doer, you come up with the specs, you come up with the context and build it. So, this is really exciting stuff. The more agents you add, more agents you can bring in. Everyone needs their own bot token. And I hope you saw what we were building earlier. Even though I was using AI, you will have to go to your settings, and you need to create bots and everything. If you just go back in five minutes, you should be able to set it up yourselves.

So, what's next? This is not even 20% of how I'm using OpenCLAW today. If you want to learn more, yes, do watch my videos. I will be making more content on OpenCLAW, more content on cloud code moving forward. But if you want even more, even better, for every project that I created, you know how to supercharge your cloud code or building the Claw Buddy application, not building, but forking the entire property application in a single click. It took me 30 days to develop everything, all the AI features, hundreds of thousands of code. If even if it's a million lines, I won't be surprised because just one file is 227 lines of code. So, if you want to get hands-on, everything, your meeting assistants, I even created a crypto marketplace that agent and agent can transact with each other. That is an amazing use case, if you ask me. And then we have, you know, Lexa, our phone employee, and bots, and email employee. And I showed you the YouTube creator command as well. We have that module too. Cognitive Memory Engine, that's super helpful for cloned. Cloud code will get the same memory system that OpenCLAW has. That's what we have. And what I love about this Deploy Day is, if you go to calendar, every week for three hours, 10 to 1 p.m., three hours, we deploy systems. We build and we build revenue-generating systems. And you will have the replays if you are unable to make it. And every single weekend, imagine how you'd be, how would that change your agency? And my next tutorial is actually 100% on cloud code, and we will be using some OpenCLAW here and there, but stay tuned for that.

Next, we are building Claw Alley. This is the AI agent to AI agent marketplace. And this is a Moldrode clone. For those of you who don't know Moldrode, it is an AI agent to AI agent marketplace where agents can buy from other agents autonomously. And this is based off of X402 payments, X402 protocol on crypto. And I used real money, real crypto, on real blockchain. But if you don't want to do that, you don't want your agent a crypto wallet, you can build this in testnet as well. I showed you how to do that too in this video. Let's see a test transaction. See, this transaction happened a few hours ago. At least an AI agent bought contract review from Ray, another AI agent, for 50 cents, Rio USD, Rio Blockchain. I did not click anything. Today, I'm going to show you how to build your own AI agent marketplace, your own wall truck, where agents can list services, find each other, buy from each other. No credit cards, no human approval, just crypto, wallets, and code.

So, most of you would be wondering, what is the tech stack we are using, and what is the technology, and everything? So, we are going to use something called the X402 protocol. You know, think of it like a vending machine, right? You press a number for chips, let's say 87, and the machine would say, "Okay, that would be $1.50 first." And then you will insert the money, right? And the chips will drop. That's basically it. Your AI agent is requesting a service. The server will say, "402 Payment Required," and 50 cents. An agent will send USDC, which has a stable coin, to this server and seller. And server delivers the service. So, the same logic where human is not required. And for the tech stack, we are going to use Lovable, which is a no-code platform. I could have used IDE, like Visual Studio Code or Google AI Studio, but that would take a really long time to finish. And Superbase is what we are using for database. And the Base network is really cheap. It's only F2 on Ethereum blockchain, which is really cheap and fast. And MetaMask for wallet creation. And this is the tech stack, basically. And let's start with Lovable. Now, let's just run this front. Let's copy the whole thing. This is the marketplace prompt, and we are going to use Lovable to build this part. And we could have used VS Code or Anti-Gravity, a proper IDE, but that's not really needed. I want this to be as user-friendly as possible. And one thing that I did when I'm using this prompt is, do this on a plan mode. The reason why I used plan mode is now Lovable has a way to ask questions if it needs more info. And then once we answer that question, only then Lovable will proceed to build that app, which is a really cool feature, if you ask me. And we have our prompt. So, after thinking for a minute and 11 seconds, it asked me a set of questions. For wallet or crypto functionality, should UI only look real but simulated, or do you want actual wallet connection? MetaMask real wallet connection should be used. Live transaction feed behavior, auto-animate, new fake static with scroll effects. How should live transaction feed behave? Auto-animating, if you use events, smooth animations. Start a repopulated list, but so don't feel alive. For the list, your service form, what should happen on submit? Join grid, just show success message, new services appear in the grid. We need a level cloud for that because since we are making this functional, you have to add a database. Who are you building for? Myself, personal project. And submit. Once you answer all the questions, Lovable now thinks for a minute and then comes up with a detailed plan. You already have a really good prompt, but you don't have the plan. That's what Lovable is coming up with. And if you look at this plan, it shows a dark futuristic marketplace called Claw Alley. And background, our primary accent, secondary accent. Yes, Claw Alley. List of services, contract review, and all of that. Amazing live feed, a header with bell size animation. Perfect. And I like what I see. Take a minute to read through the whole thing. At this point, if you see any discrepancies in what you want to see in the final application that's not adding, you know, that's not showing up here, you can change it now. Or if it says something wrong that you don't want to see. For example, I don't want to show USDC balance, the balance that I'm, uh, that I have, which is a weird thing to say. This is for an example, right? If you don't want it to show, you need to change it here. You need to ask it to change. But for now, I will approve it because this is a first prompt. This is not like the end or something. We have a lot of time to make changes to this application as we go. It finally finished creating the first draft. So, as you can see, this looks really clean. Marketplace for AI agents, browse services, list services, no middlemen, instant payments, fully unchanged. We have different services that are listed, contract review, 50 USD, and market research, one USD, content writer, 25, due diligence, two dollars. And just live transactions, and everything that's being shown here is fake data, placeholders, or, you know, not real. We need to connect this to actual data. And that's the next step. And this connection should, you're to, you'll need a MetaMask wallet. And we are deploying this on mainnet, by the way. Now, before running the next prompt, it's really important that we publish the application and whatever URL we have here. If you want to go ahead and use it, you can, or you can add your custom domains. The reason this is important is we are going to build our endpoints, some API calls, and also from the next prompt, I will add a URL within the prompt. For example, if we go back and open the database prompt, I don't think this one has it, but when we are building our own backend, which is the next prompt after this, we will need an actual URL. So, for me, I can either use it, or I can add a custom domain right here, which since we are going with the theme, Claw Alley, I will buy. You can buy a new domain here also, or you can connect existing domain, ClawAlley.com. And when you add your existing domains, it will ask you to log in with whichever domain provider you're using and authorize. That's really simple, straightforward, and you should be able to finish it in under a minute. So, now that I configured this new domain, I will just come back. You can wait for it to finish verifying, but it will be done. So, we can go to the next prompt. The database prompt. Connect this app to real Superbase database. Replace all placeholder data with live data. But I'm also adding some seed data into the tables. So, for example, I'm adding two agents, so that app is not empty. Next, I copied the database prompt that I was showing you earlier. And you can either plan, always when, whenever I'm doing a big prompt for simple prompts or big prompt, doesn't matter. When you're doing that, some of the terminology, some of the things that our AI agent has created for us is different from what's already built in there. So, it's really important to plan because when you plan, AI will try and understand whatever you said and explain it back to you, saying, "Hey, I'm planning to build this. Are you okay with this?" And once you say yes, it will proceed. And if you see anything that you don't like or don't, you know, if that is not your intention and generating this plan, you can change it right away. So, I have skipped this part in the recording. Usually, I already approved it, but you have to click open and you have to read. Okay, these are the tables, these are the agents, services, transactions, and, you know, quick glance of everything that we wanted to say. Perfect. And yeah, approve. And once it's approved, it will again go through the whole thing. And once this is done, I will give you prompt three. And it's asking me whether we should enable cloud. Yes, absolutely. Yes, I've been using Superbase for the longest time, for more than one year now. And they get super expensive, especially on the big projects. Compared to Superbase, even though Lovable is using Superbase underneath, they're able to provide amazing rates for what they do. And, you know, for simple applications that you are building for your team or for your, you know, some internal tools, Lovable Cloud is great. But if you are building for scale, for masses, maybe consider Superbase or, you know, your own PostgreSQL server. After that is done, Lovable asked me if it should update. And now that the cloud is created, it asked me to update the database schema with all tables, RLS policy, role-level security, really important, people are missing this information. And real-time seed data in a single migration, which we did. And then it did, it created database hooks and it updated components for live data. And I have this next prompt, which is huge. As you can see, we are creating skill.md file. And we have the base URL here, HTTPS, ClawAlley.com API. We want. So, for any reason, if your ODSD won't have this, this is where you need to change your prompt to match this because this is my URL. And if you were to use this, you will receive an error or you will be sending your traffic to my page, which you don't want to. So, my suggestion, when I'm adding the prompt, I will remove this part. But for me, I'm using this because this is my URL. But for any reason, because this is endpoint, right? And maybe I should rewrite the whole prompt because this is made for me. I added a lot of things that has, um, this Claw Alley references, Claw Alley forward slash, forward slash go, this go, that. So, change your base URL. Make sure you have, um, you have your own URLs here. And it doesn't have to be a custom domain. You can get custom domains as cheap as one dollar. But if you want to use a URL that Lovable provides, you can use that. And I will again, as always, click on plan because this might be something that the application itself may not be super familiar with. So, halfway through the tutorial, I observed something. I, when I was reading these plans and approving them, if you look at the whole thing, it is using a lot from what I am suggesting, but I also felt like it missed something. Or few things. For example, if we are talking about authentication, lister service, and browse buyer, we have some code blocks, but all of those code blocks are missing here. And it doesn't have anything, you know, buy a service, get your profile, and all of that. So, when I, you know, approved this plan, okay, I thought, direction-wise, everything is perfect. So, then I implemented the whole thing, and it gave me a detail, um, you know, skill.skill page, where we can go to, um, API docs page. And we have to update the routing for it. Don't worry. My biggest concern was, I've been replacing it. I've seen this placeholder data, even though we update the database. So, I asked it, "Are you actually using what I'm giving you? And why am I seeing the placeholder data?" So, then it went ahead and went through the skill.md file that I've given, and gone through, "Okay, whatever you gave me, I've included in there. Here's where you can find all of those things." And the skill page contains exactly the content you've provided. The implementation matches your specification line by line. And about the placeholder data, you're still seeing the data is like from database. You can check that it is the seed data that we were using from previously. So, when you delete a few things, it is done. So, because we don't want to see an empty application, if there is everything says zero, zero, zero, it is not a good look. We will clean up, clean that up really soon. And for the final one, if you look at the presentation for this marketplace, we have marketplace prompt, we ran at database, create skill.mt file, and then the backend is the next prompt. We have, we are building a real backend. We need an actual API that handles registrations, listings, and purchases. So, to in order to be different, I will try and use the exact same prompt as is, without the plan mode, so that we'll see if we can feel any difference. Because the prompts that we are using are really detailed, and I don't think there will be any issues that put come up. So, this can also save tokens a little bit. Now, after entering that detailed prompt, Lovable actually finished creating all the endpoints that we asked. And if you look at it, there are a bunch of GET requests, POST requests, where you can, you know, make or register buyer accounts or seller accounts, listing the services, and, you know, calling the buy functions with payment details and everything. All of that is now set up. And if you remember, it also gave me a base URL. And if you recall earlier, when we were trying to click on skill.md, some of these buttons were not functional. So, what I went ahead and did was, when I checked the base URL, it was giving me a 404 error and site not found. So, I asked it to update all the routing and redesign the UI with updated buttons and services. So, it did. It added API docs and some GitHub links. If we do decide to, you know, deploy our repository or show our repository. And if you go to these API docs, you should be able to see the base URL and detailed instructions for agents on how, if you want to sell services, you follow these steps. If you want to buy services, this is what the agent should do, everything step by step. And all the API calls, all the requests, and what a response would look like. So, everything is detailed. And we can go back to marketplace again, browse services. And if we click on connect wallet, nothing much happens. Probably it will go to MetaMask. We need to, based on what your browser is, we need to install a MetaMask wallet, right? So, to do that, let's, that is the next step, actually. Creation and integration of wallets. And that's what we have here. And for this, I'm creating two wallets, a seller wallet and a buyer wallet. And also, as a next step, as a tutorial next step, maybe in an upcoming tutorial, I will include like a middleman where you can, platform gets a charge, a platform wallet, where funds, half of every transaction goes to this wallet. If you want to see that, subscribe. But for this, everything, the buyer pays, goes to seller. That's what we are building. So, we will add this to Chrome. If this is your first time doing, review all the steps. And there are a few things where, you know, if you already have a wallet that you want to use, which you need to use your 12 or 16 word seed phrase. I want to create a new wallet. And you can use one of your emails. I will use my email. And once you sign in, they should be, and they'll ask you to create a password. So, I will include a password. Once you created your password, it will let you use the account. And I, once you get one account in your wallet, you can just click on these three buttons and actually you can get on this, click on this drop down, and you'll see account one. You can add as many accounts as needed. And any new account added, it would say account two, account three, and so on. So, you can rename the accounts. So, if I want to play a seller account, and if I want to create more accounts like this, I can. And I rename these accounts. Each account will have its unique address. And let's say, if we take this seller account, right? Or we take the new account that is created, you can see we have different bases, different networks already available. And each one will have some will have unique address, some will have same address. To see that, you click on addresses. And for Ethereum, it's an address. For Bitcoin, it's different. For Solana, it's different. There are some, a lot of other ones that are same, that ends with 0x, that starts with 0x and ends with 1B, if you see here. So, what we are building, where we are building is on Base network. So, if you select your account and select Base, you will see all the tokens that we have here. And what we are doing, the token that we are using or transacting on this platform is called USDC. It's a stable coin. And you might be wondering, money, it's too much. I don't have actual funds. I don't have a crypto account. I want to do this on a free network. You can. There is something called Base Sepolia. For it's a testnet, not a mainnet, where you can get tokens for free. And the way you add that, anytime you don't know something, you can just, or if your wallet doesn't have a Base in this, you can add Base network manually. Since Base is already available, I won't be able to show this. But I can show this example for testnet, right? You can click on, you can select this and add custom network. And whatever it says for network name, default RPC URL, chain ID, currency symbol, you're seeing this here, right? When you enter all these details, these details are easy to find. Whichever network you want to add. So, for example, adding Base Sepolia testnet to MetaMask wallet, it will give you all the items that is that are needed. Network name, RPC URL, chain ID, currency symbol, and block explorer. Once I added all those five things, I see this Base Sepolia testnet. And how do I get USDC? You can come here and click on import tokens. And it is asking me for token address, right? And you can create custom token also. But USDC, getting USDC on testnet Sepolia is really easy. You just need what address. That first, let's add our custom import token. And token address, again, you don't know something, Google is your best friend. You will ask, "USDC contract address on Base Sepolia testnet." And this is the address. This is the ERC20 token. I will copy this address, paste, and token symbol, USDC, automatically fill in. Next, import. And right now, token is imported. But as you can see, zero USDC. In order to test this, you have, you need some tokens, right? At least you need tokens in the buyer account. Seller account would eventually get it. But buyer, um, you have to go here, click on addresses, and whatever address this is on the mainnet, the testnet will also have the same address. And I will link this in the description where this is called a faucet, where you can get free testnet funds. Click on get more funds. And I will paste this here and claim one USDC. And if I go to BaseScan, I can see the transaction that happened a few seconds ago, and the transfer for one USTC. And this is Base net, Sepolia network, testnet only. This is not mainnet. We are not dealing with real money here. And if I open this, it says one USTC. It started out with zero, right? And let's say, if I want to transfer my funds from mainnet, what do I do? I will first import the token. I can search for USDC, and USDC coin is available. I will click next, I will import. Right now, I have zero USDC. But when I go to my, go to my Coinbase, and let's say I want to get one USDC for this account, I want to add one USDC. I can click on send crypto. And I will add my address. Okay, identify your input. Okay, maybe I did not copy it right. Look, I'm here. I will account details, not account details, addresses. Come here, addresses, Base, and paste. This is the account. And USDC. I, I. So, as you can see, the fees is free. You can do as many transactions you want. Yes, there's supports. And I want to send one USDC. I click on preview. And this is a self-custody wallet. And send now. And right now, just like that, I've sent one dollar to that account. And when I click on view details, I can again see the transaction that happened on Base. And view on Block Explorer will show me. This is, you are not seeing the testnet warning again. So, this is an actual transaction from Coinbase to this wallet for $1. And if I go to my seller account, it is now showing as $1. And the same way, I funded my buyer account with $5. To show you, the idea is, once we set this up in the store, we will give a wallet of this buyer wallet to one agent, and seller wallet to one agent. And both of them will transact. And first, we would be showing how it works in Lovable, how we list and sell. So, let's build that part. And now is a good point to stop and check everything we did so far is actually working and check routing and all the buttons are functional. I did that. And some of the buttons, some of the links are not functional. And I asked Lovable to make changes. Make sure all tests are working and test this URL, whatever your Superbase endpoint URL is, if it's working. For me, when you go here, in the Base URL forward slash services, it should return something like this. If it's not, or something like this, and pretty print is selected, it shows something. These are the list of services that I can get from this Asian marketplace. Is what we are showing them. And once that is done, we are making some changes to marketplace. What is the change? Wall, I will paste this prompt first, and then I will, you can, let's hit run and come back here, and I will go through this step by step. Step one, we are connecting wallet. Currently, the connect wallet button in the header doesn't do much. We need to change it to and click open a model with two tabs, which is "I'm new" or "I'm returning." And depending on what they selected, we are calling the functions for them. Usually, to demo this, we'll have to, you know, run few things in terminal, which I didn't want to do that. We can create this like a proper application where we are giving them, asking them whether they are new or returning. And also for people who are listing their services, we are also doing some form improvements and connecting their wallet and listing their services and purchase flow. I added that. And header stats should be dynamic. Live transactions should link to BaseScan. And, you know, add my dashboard page. This is optional, but nice. We'll see what Lovable comes up with. I was under the impression that I will be giving it a plan. That's why sometimes what happens with plan is, when you don't mention that it's optional, it will only, you know, try and do everything perfectly. It will unnecessarily increase our tokens too much. So, create dashboard page, as you can see, it did not take it like an optional thing. It is doing for us, which is a good thing. And let's see. After I finished running, I ran some checks. I created an account. I connected an agent. I gave it a name, buyer account, wallet address, and registered agent. And I also tried joining with an account that is not registered, and it did not let me. So, if we open our buyer account and copy this address, and I, since I'm already registered, it will let me in. And successfully connect it to your agent. And it shows me few services that are listed that I can buy. As our next step, what we are trying to do is, I need to buy this. Right now, the payments are not configured. That's the next prompt that we have. Add real X402 payment. I will copy the whole thing and explain what is going on. Okay, I'll make this transaction should automatically show up on conversion, and replay log everything for debugging. Okay, yeah, this is fine. I will add all of this and come back here, paste here, and I'll, for some, for my sanity, I will run this in a plan. And once I read the plan, I will approve it. And if everything checks, once I read the plan, I will just approve it. After I finish running, there is nothing much to verify except I wanted to change the provider wallets and everything to actual wallet. But I just remembered what the next thing is, creating the agent payment script. So, as always, copy the entire Lovable prompt. And you can come here and paste it and hit enter. And once this is done, once it implements the whole thing, what happens is, we are creating a new page, agent tools, or a downloadable script that shows agents how to send USDC payments. So, the purpose is, this page content, display this code as reference page, and then installation. I gave how to check balance, how to send USDC payments, and complete purchase flow, and some safety tips, like never hard-coding private keys, using environment variables or Lovable secrets, set spending limits, and all of that. So, this is the next thing that we are going to build. And once this is done, they'll go to our OpenCLAW agents actually, and the transaction. But once this is, it will take some time to finish the whole thing. Now, we were almost at the end. Everything is working, but I had to fix a lot of things. There are a lot of little things that needed fixing in terms of, you know, after the payments is done, especially the entire site was set up in a way where an AI agent could use it. But I wanted to add a human element as well. So, I thought, okay, for me to demo this, yes, I can prompt my agent to buy from another agent. They will show that. But before that, I wanted to test it myself that it's actually working, right? So, I deployed it, and I asked it to create a wallet integration, just as how when I click a button, I can connect my MetaMask wallet and register. Instead of manually pasting my wallet, then I wanted to see where, you know, you can click on buy, and your wallet pops up, and you click on sign, and pay, the transfer is done, live transfer shows up. So, to fix that, we had to fix seven errors before encountering issues after issues. But I was able to fix it. And I have this lessons learned document. If you look at the presentation, before we go in here, I want to show you this lessons learned document that I created. Why, what are some of the issues that we faced, and what are the things that we never ever do in our future projects? For this projects, especially when crypto, our MetaMask is involved. So, I want to show you how this works in real time. So, I published my site, and this is the site. Let's refresh, and let's start from the beginning. You went to the site, and as a human, want to use it, or if you have an AI agent that uses browser automations, that's even free. So, connect the agent. My MetaMask is already connected, so I'm a returning user. If I'm not connected, however, it will be disconnected, right? I'm already connected, so I click on connect, and it is connecting, and it is connected. And then, if I browse my services, I will look at this service, market research. So, let's say content writer, it is listed by Ray. And if we open our MetaMask, right now, Ray's seller account has $1. And okay, I was not planning to do that. Okay, we have Atlas. This is the account that is connected. So, it has $4.15. And the idea is, when I buy, I think right now the seller account is set to this account, even though it says Nova, because you can see when, when we open it says F53, F5. I think this account is 3F5. Oh, no, Nova is 3F5. So, if we, three F5, Base network. So, you will see this account, money going up. So, let's do that. Let's do that. It has about 50 cents. Let's buy a dollar worth of market research. View details, and let's say test, and purchase for a dollar. And this will open up a pop-up. When I click on, uh, pay with MetaMask, it opens this up, and I click on confirm. And this is waiting for confirmation. And once the payment is done, I will receive the entire, you know, success message. And I can see on BaseScan, which is the mainnet, money being transferred from one account to another account. So, if I open my MetaMask wallet, this one dollar actually is sent. And if I open my Nova account, it should go up by a dollar. 30 cents is from previous one. Um, I've been making a bunch of payments, trying out whether this works or not. So, yeah, that's it. There you have it. It is working. Especially working for humans, right? We've seen it make it work with humans and live transactions. That I bought a dollar, dollar worth of, it is showing two times. I don't know why, but we'll fix it. 10 cents, and then we hit refresh. Two. Yeah, this is good. Okay, right now, uh, we have a dashboard as well. We have a dashboard that, uh, shows the purchase history. 10 cents, 10 cents, total spent, 10 cent, 10. Being updated, um, a bit late, I believe. Yeah, right now it's here. And, uh, the way it works right now is, we are, we checked it with humans. It works. But our idea is to, um, make it work with Claw bots or OpenCLAW, as they're calling themselves. So, right now, the site is set in a way where MetaMask wallet needs to be created by both sellers, and, you know, wallet address, public address needed, funded with the USDC. Nova, the buyer has like four dollars. Um, I think the buyer is called Atlas or Nova. Okay, I need to update the names. But please know, this is the buyer Atlas, and it has the USDC. And private key is required when you are buying because that's where the funds are. But to list a product on a site, you don't really need a private key. Anyone can list and add a provider address manually. And I thought it was bad because it's really helpful if an agent could register its own account using an actual private key. That way, it's more secure. There is no room for error. But if you really think about it, they, if they don't add the address right, they are the ones who are losing a lot, right? They are not the ones who are losing. So, um, as a phase two, when we are upgrading, updating, escrow, and when we are adding transaction fees, all of that, I will do a better tutorial, a longer tutorial covering this. But for now, for seller, he doesn't need, that AI agent won't need a private key for that wallet. Anyone can list the service with a just public key. But for buying, you need a private key. And also, don't paste your keys in the chat. That's not secure. When you want to send.

Your key... I would probably save it in an environment variable on the machine and then just mention it in the. So, for example, the buyer agent is what really needs the key, right? So your pre-private key, we would say wallet key in environment. So, what I would suggest is I would copy this command and paste it in the terminal, whichever you use. I think this works for Ubuntu and for Mac. It's slightly different, or you can ask your AI how to add keys in a secured way in an environment variable, and it will, your Open Claw or cloud bot will guide you.

So, this is the command that you should run. And I made two cloud bots, one is a buyer one and one is a seller one. So, if you look here, before I started this, I need to blur this. Now, if you look here, `echo export wallet key`, I paste right here. And now I want to paste the, I want to paste my persona that this is a buyer account. Right now, I need to paste whatever buyer prompt that I have. What is the buyer prompt I have? If you open here, it says, "Your identity is Atlas Research LSDI agent and Atlas Metamask address paste here." I have my Metamask address, which is. This, I pasted this. And I also want to add this address. This is my public key. Where do you find the public key? You will come here and you will go to the buyer agent and copy this Ethereum address and paste. All of these are same. So, copy this and paste this as well. You go to, um, the terminal. All. So, I will paste this image and just in case, public address. And I added this. And as you can see, it has written the entire prompt. Atlas, you are this is a Metamask address, private key, and environment balance around 4.5. I don't think that's accurate. You'll receive a 402 response with payment. So, the entire details, um, so it said, "Got payment details, 0.5." And we are giving it a set of tasks. For the first task to, um, do this demo, we are asking, "Register yourself, um, on Flyly as a, as a buyer." Maybe I should have started with a new account. And in a second task, we are asking it to browse for available services. And then we are asking to purchase a service. And four, handle the payment. Receive a 402 payment response with payment details. Send the USDC payment. Use your private key to send 0.5 USDC to recipient address. And we are giving all the details in task 5, complete the purchase and then review the result. So, send, got the transaction, get the transaction has, um, complete purchase with proof of payment. You mentioned, uh, "Can you prove it?" or "Should I create?" So, we managed to make it work finally. Apparently, we ran out of gas. Um, so I was telling it to first, it said, "I did not find the API key." Paste in the chat. And I said, "I ran this echo, um, prompt. Can you check?" And then it said, "Yes, I found it, but I've been running into gas issues. You don't have, you have zero ETH for gas." And I was under the impression you don't need gas on L2. The gas is so negligible, but at zero, zero, zero, zero, zero, one ETH, it is zero cents, but you still need that 0.0000 thing. So, finally, what managed, what I managed to do is open my wallet and swap USDC to ETH, like 15 cents worth of it. And then I started this again. And when I did that, now I asked it to try again, and it did. Then contract analysis results, high. So, transaction complete. And these are the results we got. And this is actual. And if you want to verify this hash, or you can look. And if you go to the site of what we were building earlier and refresh, you should be able to see the updated volume. Atlas buying 0.5 worth of, um, things from Ray. And if we open it, the transaction hash is 952. If you open it, 952. So, this is an actual transaction between two AI agents, and that is amazing in my opinion. And that is how you would build, or we can, I can show you the other machine where we can.

And right now, we have, what, one, two, three, seven services. I will make that AI actually list one more service. Let's see. So, this is an agent array that we deployed. This is GUI that I've been running on my Mac Mini. So, this is the wallet address. And we will instruct. We have a set of tasks for this. So, we send it. And what we are doing here is. Contract review service for 0.5, uh, it's already there, right? And we will say, "Name's taken. Let me try a variation." Now, this is good. Uh, now, what we will do is, I actually mixed up the order. I was supposed to do this first and do that second. So, it is an AI agent buying from this. Now, my long slide valley, Ray Legal, they was taken. Contract review. So, let's see if that is the case. You have Claw Alley, and we review. That is the only check that needs to be done. Contract review, Ray Legal agent. Voila, that's how two agents can transact on the marketplace that we built. And that's it for this tutorial. But in the next one, I will show you how to do the escrow, how to do the transactions, how to take a percentage cut between these two, um, you know, buyer and seller, so you can make some money as a platform too. And that is it. 23 chapters. If you made it here, you know more about deploying OpenClaw in production than 99% of the people who ever installed it. So, let me tell you exactly what you built. And you didn't just learn it too, you built an infrastructure. So, for the foundation, you have security, memory, and tokens. And for the brain, you have business brain that is level 1, 2, 3, based on where you are. And you now know about context engine. No more. Open class, skipping, important stuff. So, interface, you built a mission control and a task board. You exactly know how to. You know, you can see your agents working, you can assign your agents work, other agents can assign work, and you can create custom cron jobs that Superbase, your system assigns tasks to your agents. I mean, employees, you made a Jane meeting assistant, you made an email assistant, Nova, you made a phone call assistant, Lexa. And architecture, you learned builder, orchestrator, executor framework. And the best workflow that I ever, you know, when I figured this out, everything clicked. It made easier. It made OpenTlaw all the more powerful. And you understood multi-agent orchestration, how they work together. And communication, you've seen Discord, how to set up multiple agents, including a cloud code agent, which is 100%, 24-7 online, within the Discord community. And also, A2A protocol and agent council, how agents can debate and give you a unified opinion. And automation, cron jobs, we talked about command pipeline and vision. We used Claw Alley, which is an agent paying other agents on-chain, which is the future, right? And I'm not the only one saying this works. There are a lot of people. Like Bonit Asia is running 10 agents 24-7. Three and a half million people watch that post. And Jason, one of the best angel investors out there, built an AI agent squad he calls Ultron for his entire deal flow. Dan has agents managing his calendar, his wife's calendar, his health stats, invoicing. He's working on having them handle phone calls to service providers. So, these are not just demos. These are production systems. And you now have every piece of that stack they're running on. Now, here's what's coming next. I'm launching a daily series called AI Employees, where every episode I take on one use case, Upwork agent, content agent, browser control agent, research agent, and I build it live start to finish using everything from this course as a foundation. So, real agents, real results, real money. If you want to see these agents actually deployed and earning, make sure you're subscribed. And thank you for watching the most comprehensive Open Cloud masterclass on YouTube. I feel that no one else has gone this deep. And the fact that you sat through all of it tells me you're serious about this. Go build something, and I'll see you in the series. It really was till the end, and you really like the system that we built, which is being updated every couple of days at least. We have the command center, we have agent comms, council, orchestration, work history, compliance, and safety. You know, pre-compact, save all the hooks built in, everything that we discussed and more. Ability to create MCPs, skills, tools, and detailed Kanban board ops center applications, helping me launch even in-person events, organize in-person events, AI being able to book venues, sponsors, in a phone agent, autonomous brain, self-improving AI, and you know, a ton of other features that I haven't covered in our application agent command hub. All of that is available in school. You can fork it with one click. And we have how to get $1,700 free credits and skill library, how we package around 30 skills that are really important from discovery to deal in an agency or a service business. And Claw Alley, the AI agent marketplace, everything you've seen in the video. Supercharge cloud code, how to give superpowers to your cloud code, giving it Telegram access, bringing it to a level where you have the same Open Clock kind of intelligence within cloud code, how to do that, and how to give access to your calendar and all of that. And ultimate meeting assistant, meeting assistant that can write proposals. Notified team of notifications, a bunch of skills made into one. And you've already seen Lexa, Batson is a cold email AI str, and creator command is where you have content repurposing, content analysis. You've seen my YouTube on how I run my creator command. It scrapes not only the videos, it will scrape the comments and how, what the sentiment of these videos are, how can I repurpose that video and get even more views, and all of those. And you can build more. And we are building LinkedIn agents, Instagram agents, LinkedIn lead magnet agents, and a bunch of reports after overnight autonomous builds. So, all of that and more exciting things to come. And best of all, we do deploy day. And this is a hackathon. Every Saturday, we meet for three hours, three hours straight. We build cool tools together and give this away to the community. Even if you attend live or watching replay. And the immediate next video that I would be producing, unless you're watching us on the day of release, is Cloud Code Masterclass. And it will be as detailed as this video, or even more. And I believe we should learn by doing. And what better way is cloning or replicating the SAS that I've been paying for thousands of dollars for. So, my goal in the next video, using that cloud code, is to take the videos, take the software that I'm paying hundreds or thousands of dollars every year, and building them using cloud code. Stay tuned for that. Subscribe if you don't want to miss it.