Transcription
When you first hear about OpenClaw, it sounds like magic. You get told it's a powerful AI employee that can run its own computer, use your apps, and work while you sleep. But when you look into the concepts, the gateway sold out MD heartbeats skills, it's quite a lot, and you don't know where to start. And that's completely normal.
If you're new here, my name is Jay, and after spending hundreds of hours with OpenClaw, building real systems, running our AI solutions practice, and managing one of the largest AI communities globally, today I'm giving you the shortcut I wish someone gave me on day one. These are 20 plus open claw concepts in plain English explained in more or less 60 seconds each. Let's get started.
[music]
So first of all, what is openclaw? Well, you already know tools like chat GPD and claude, right? You type something in those apps and they reply to you, but that's all they do. They talk. They can't manage your calendar. They can't post and manage your social media accounts for you. And they can't manage their own email inbox just like a human would. Now, OpenCloud can do all three. And that is just the start. So imagine you're upgrading from texting a smart friend to actually hiring a full-time personal assistant. And as you'll see later, the reason why OpenCloud can do this is because it has its own dedicated computer. So it's able to take action. It can browse the web, manage files, and even connect to the applications that you use. If Chat GPT is a chatbot that you can just converse with, then OpenClaw is sort of like an employee who never clocks out. It's also completely open source, so anyone can get it. And it's grown faster than almost any other software project in history. And because it's fully open source, no company can shut it down, pull it, or lock it behind a payw wall.
So, how do you actually get it? Well, it's actually surprisingly simple. If you go to openclaw.ai and scroll down to find this oneliner command. If you're on Windows, this is the command that you should copy. Or if you are on Mac OS or Linux, then this is the command that you should use. And to use that command, you just need to open the terminal app in the computer that you're installing OpenClaw to. And if you paste in that command, click on enter. that will now run through the installation process for OpenClaw in that device.
Now, what's important to understand is that OpenClaw runs as a background process for that computer that you're installing it to. And it can also fully create, edit, use, and even delete files in that computer. That's why right now it's best practice to just install OpenClaw on an old computer that you may have lying around or some people in the community also buy a Mac Mini specifically just for this software. Or as an alternative, if you don't have a spare laptop or you don't want to buy your own Mac Mini, then what you can do is rent out this VPS, a virtual private server, basically a computer in the cloud, which if you do a quick Google search, you'll find providers like Hner, Digital Ocean, and Hostinger, where you can basically rent out their computers for a few dollars a month. That way, your agent is always online, even when your home computer is off. But the key takeaway here is that you should treat OpenClaw like a new employee hire. And so instead of giving them access to your own laptop and your own device, it is best practice to get them their own computer, their own workspace, so that whatever files that they create and edit lives just within that environment.
And by the way, if you're interested in going from just using AI to getting paid for it, then check out the Robo Nuggets community down in the description. We've got founders in there who landed their first client in weeks, live build sessions where we create this stuff together and the actual templates behind what I just showed in this video. The community is also the reason these lessons get made. So see that below if that's for you.
Now an important concept to understand is the way that you connect Open Claw to your AI models because this has a lot of implication especially around the cost because if you don't set this up properly open claw can get quite expensive. Now when you install openclaw you will now have a choice generally between two setups. One is to set it up via an API key basically an application programming interface key or via OOTH which stands for open authorization. Now, without getting too technical between these two, the important distinction between them is that when you use an API key, you pay per usage. And when you do that, there is no ceiling to the amount of token cost that you can incur this month. And so, generally, using an API key is a bit more advanced and is mainly geared towards developers anyway. And so, if you're new to this space, generally what I would recommend is to try and set up an OOTH. And the benefit to that is that you incur a flat monthly fee. Let's say you subscribe to ChatGBT Plus at $20 a month, then you'll only pay for $20 a month regardless of how many tokens you use.
Now, the tricky thing to understand when it comes to OOTH is that since this technology is so new, a lot of the top model providers, namely OpenAI, Entropic, and Google, some of them actually have unclear stances or statements whether it's okay to use OOTH for OpenClaw or not. And the reason why this is is because OOTH is pretty subsidized, meaning you can actually use a lot of tokens for cheaper versus if you go the API key route, which is more expensive for you, but obviously more profitable for them because it is uncapped. Now, the good news is Open AI since they just hired the creator of OpenClaw, they actually came out and said that OOTH is allowed. And so, if you're already subscribed to the Chad GPT plus or pro plan, then just use that. Now, if you want to use Entropic and their Opus 4.6 model, which is really popular right now, this is actually a gray area. Some people claim that they've been blocked, but there's actually no official public statement yet. And I still know some users who are still using their OOTH for Open Claw with Entropic, and it still works as intended. Now, with Google, as of the time of this recording, there have been several documented instances already of their Gmail accounts being banned from this usage. So, I think this is probably the most risky of the lot. And so, if you're using OpenClaw for the first time, then it does seem that currently OpenAI's OOTH would be the way to go.
Now, an important concept when it comes to AI agents like OpenClaw is the agentic loop. This is the one that separates an agent from a chatbot in basically how AI agents work today. Because when you talk to Chat GBD in a browser, what happens is you send a message, it sends you a reply, and then that's done. It's one turn. But an agent works in a loop. You give it a task and it figures out the steps on its own. It calls a tool, it maybe reads the result, decides what to do next, and then calls another tool and keeps going until the job is done. So, all in one turn without you having to do anything. To provide a real example, let's say you ask it to fix a bug. What it will do is read the file, find the error, edit the code, run the test, sees it fail, reads the error message, and then tries a different fix until it passes. And so you ask one question, and basically in that whole loop, it took six actions. Every concept in this video builds on this agentic loop concept. And so the memory, the skills, they all exist to feed this loop.
So once you have OpenClaw installed, where does OpenClaw actually live? Well, it runs as a background process on your machine, right? So you can think of it as an engine that is always on. This engine is called the gateway. It's sort of like the receptionist at the front desk. So let's say you message openclaw through your WhatsApp, your telegram or your slack. The gateway figures out which conversation they belong to, loads the relevant contexts and passes them to the actual AI models underneath. And for every OpenClaw install, you also get a web interface that you can open in your browser. and this web interface if in case you don't know where to find it what you can do is just ask your open claw to send the link to the gateway web interface and it will guide you on how to view this now personally I don't often go to this gateway dashboard because it has a lot of information that isn't really something that you need to see for daily work however this agents view is quite relevant so that you can understand how your agents are built which we'll show in the concepts that we'll be discussing next so the gateway is the engine that handles everything internally but how does openclaw actually reach you and talk to you well this is done through chat channels and channels you can imagine as phone lines basically plugged into the switchboard which is the gateway in this case and these channels connect your openclaw to a different messaging platform. So if you want WhatsApp, Telegram, Discord or Slack then that is something that you can do. You set them up once and from that point on you can message your AI assistant the same way that you'd message a workmate. So if you're on mobile and you have an idea in mind then you can just message your OpenClaw agent via Telegram and that will route through the gateway and comes back through the same channel. So the core idea is this. Your OpenClaw agent basically has one brain but many ears and many ways to communicate with it.
So now let's talk about OpenClaw's multi-agent and sub agent capabilities. Because as you may have noticed from what I'm showing, OpenClaw does support running multiple agents inside a single gateway engine. Each of these has their own workspace, their own memory, even their own name and their own personality if you want. So you can have a team of agents that have access to specific tools, specific skills so that you can better manage the context for each of them and their token usage, but also so that you can continue to refine how you are working with this team of agents that you have. So, for example, you might have a personal assistant agent that handles your calendar and messages, a developer agent that manages a lot of bug fixing and coding, a sales agent that drafts outreach emails, and while they all live in the same building, which is the gateway engine, they would each have their own office and their own filing cabinet, which would be their workspace. And for you to set up these multiple agents, all you need to do is to request that from your main agent, and it will guide you on how to set those up. The general principle for it to be as simple as possible is for you to just identify different buckets of work that you do and think of a role that can be filled instead of a specific task to be done. And to start with, you can begin with two to three roles that you can fill in and create agents specific to those.
Now, sub aents are actually pretty simple. When your agent needs help with something, what it can do is it can spin up a sub aent. Think of it like handing a task to an intern, for example. Your agent gives them just enough context and the sub agent goes off and works on it independently and come backs to your agent with an answer. Meanwhile, your main agent keeps working and you can talk to it in parallel. And so sub aents are just a way for your agents to delegate tasks.
Now when OpenClaw starts up, it needs a home base, somewhere to keep its instructions, its memory and its configuration. That is called the workspace. And the workspace is literally just a folder on the machine or the computer that you installed OpenClaw on. By default, it lives on OpenClaw/ aents, but depending on when you install OpenClaw, this may be different for you. So, the best way to find where your workspace lives if you want to check it is to just ask your OpenClaw agent. But the key here in this new paradigm is that none of this is code. Like, you can literally open any of these as a text file and read and understand it like you would a book or an article. It's all just written in plain English. And that is the code that sits within your AI agent's brain.
So now for the next following concepts, let's go through some of those essential markdown files that constitute the whole of your AI agent. One of them is soul.md. So if you think of these AI agents as digital employees, then one important aspect of them is their personality. Maybe you want your agent to be a bit more formal, maybe some are a bit more casual, some are a bit more blunt. And so it's the same with your OpenClaw agent. And this file called soul.md is where you define that personality. So, if you want to check the soul.md of your agents, what you can do is to go back to the gateway web interface and under agents, select the agent that you're interested in. So, you can see I have a couple of them set up already. And for my main one here, if you go to files, you'll be able to find the core files that we're about to discuss. And here you can see exactly what is in their markdown files, which essentially comprise the brain of your AI agent. So, what you can do is preview the soul.md of your agent here. And like I mentioned before, what you'll find here is just text in plain English that you can understand. I did not write this myself. This was initially written by my openclaw agent and then we just refined it as I continue to use it more. So if you look at some of the entries it made here you can see it has guidelines so that its writing sounds more natural which this whole piece of markdown file it just continues to refine as I work with it daybyday.
Now just to mention it apart from soul.md there's actually a much smaller identity.md as well and this file is very short. Basically it just stores the agents name its vive as well as the emoji that it consistently uses.
Now, if soulm defines who the agent is, then agents.mmd defines how it works. And this is probably the most important file in your AI agent system because this will be your operating manual, your rules, your priorities, your boundaries that your agent will follow every time. So, for example, if you want your agent to always check your calendar before scheduling anything or never send a message without your express approval, then agents.md is where that is defined. Now to give one example with this content agent that I spun up at the very beginning. This agents.md will be very simple where you can see it has rules around never publishing or sending content without my approval. But what's important here is that this agents.md is something that you and your agents improve over time. Now, what I recommend is that within this one pager, you include a daily self-improvement loop so that at the end of each work session, let's say at the end of the day, your agent will reflect on what they learned and propose updates to its core files, including its soul.md, its agents.mmd, and the other markdown files that we'll discuss in a bit. And to do that, you can just ask your main agent, similar to what I did here, where I said, hey, can you please make sure that our agents are proactive in updating their core files? And that will set it up for you automatically.
So now if soulm is the agent's identity, agents.md is the operating manual, then user.md is all about yourself. And this file tells the agent who it's working for so that it knows your name, your time zone, how you like to be addressed, and what projects you are currently working on. And this is just a nice way for it to store your preferences over time as well. And again, you can view this in the gateway web interface if you're interested. And you'll be able to see exactly what your agent knows about yourself. One key piece that I added here which improved my communication with the agent quite a lot is that I gave it a heads up that I use voice transcription a lot which especially on mobile what I say doesn't always get captured perfectly and so having this line in user.md helps it understand what I'm saying despite that limitation and so that's what the user MD does. It makes every interaction feel personal from the very first message.
Now the fourth component is tools.mmd and this file is really simple. It's basically your agent's notebook about how to use the specific tools in your setup. Or another analogy to it is you can think of it as like sticky notes on your employees monitor. They're not official documentation, but just the practical stuff that makes things actually work in your world. So if I give an example of what's in here, you'll see my agent made notes on what texttospech provider it is currently set up with. Things like how it connects to Obsidian, which is this note-taking app that I'm using, and even some notes around known fixes, so that the next time the same problem may come up, it already knows what to do.
Now let's talk about memory and this is where openclaw starts to really feel like a personal AI assistant because most AI tools forget everything between conversations right when you close the chat most of that knowledge is gone. Now OpenClaw works around this and establishes a persistent memory structure and it stores that memory as played markdown or text files on your machine and there's basically two types. The first one are daily notes and these are files that are stored in your workspace named as the date that captured what happened for that particular day and this includes running logs of your conversations, decisions and tasks. It's basically a work diary that your employee is using. Now the second type is memory MD and this is your agent's long-term brain and this one is a curated file where the agent saves the important stuff. So it includes some preferences, key decisions and even recurring facts. So it helps it so that it never has to ask you twice. Now, when a conversation gets really long and the agents context is about to overflow, because there's always a limit depending on what model you're using, it automatically saves what matters into memory. MD before the older messages get compressed. And this also happens silently in the background, by the way. So, it's not something that you need to invoke every time. This is why the more you use OpenClaw, the more it remembers. Now, similar to the other core files, you can also preview this in your gateway web interface. But the reason why these are blurred by default is because there may be sensitive info here. Now, if you want to check the daily memory logs that your agent has been making, they are currently not surfaced here in the gateway. But if you want to view them, what you can simply do is ask your OpenClaw to bring them up for you. Or if you have a physical computer for your OpenClaw, you can just find these files under the memory folder usually.
Now, let's discuss the concept of the heartbeat. Now, most AI tools and models just sit there waiting for you to talk to them. But the heartbeat is what enables OpenClaw to be a bit more proactive. And what it is is very simple. It's essentially a periodic check. By default, it's every 30 minutes where the gateway, which you remember is just the engine, wakes the agent up and asks it if anything needs attention. The agent then reads a core file that's called heartbeat.md, which is also in your workspace. And that contains a checklist of things it should keep an eye on and decides if it needs to tell you something or even act on something. You can also co-write this checklist with them just like all of the files that we just discussed. So, if you wanted to periodically check if you have any meetings in the next hour or to look for any urgent emails or to periodically check if your agents are finished deploying whatever it is that you're trying to do, then you can include that in heartbeat. MD and every time it wakes up, it acts on whatever is relevant. Now, by default for your agents, heartbeat.md usually begins as an empty list. So, if I preview this, you can see for this new agent that there's a note to keep this file empty or only with comments to skip heartbeat API calls. And this is important because remember the heartbeat happens every 30 minutes. So if there's a long list here then that can consume several tokens given that it's a recurring task. So be very choiceful on what is in here is a key message.
Now a related concept would be cron jobs. So if the heartbeat handles routine monitoring like inbox checks notifications that are all batched into one check every 30 minutes. But what about tasks that need to happen at specific times? That is where cron jobs come in. Now, cron, which just comes from the Greek term Kronos, is basically what you would expect. It is a scheduled automation. So, if you want OpenClaw to send you a daily briefing at 7:00 a.m. or to run a website health check or a security audit, let's say every Monday, then what your OpenClaw will do is set up a cron job for that specific time. So, if Heartbeat batches multiple checks into one turn every 30 minutes, for example, then cron would handle precise isolated tasks on exact schedules. Now in your gateway web interface, you can actually find the chron jobs here, but in my view, it is actually pretty difficult to navigate. It's one thing that I did. So just have this custom view where I have my different chron jobs set up. And in one visual, I'm able to see which particular chron jobs are set so that I can make adjustments if I ever needed to.
Now this concept is a bit more technical, but it's worth understanding because it does determine how smart your agent actually feels. Because in case you don't know, every AI model, whether it's GPD 5.4 or Opus 4.6, six. All of them have a limit on how much it can see at once. And that is basically the context window. It's measured in tokens. And for our purposes, you can just think of one token as one short word. Opus 4.6, for example, has a maximum context window of 1 million tokens. And if you want to imagine how big that is, then the Christian Bible is roughly almost exactly 1 million tokens in length. And so what that means is that Opus 4.6 can see at every one time a conversation as long as that book. And so when the conversation gets too long, the context engine is the one that compresses the older parts, basically summarizing them so that the key information survives even as the details get trimmed.
Now, a very important takeaway with a context engine is that if you remember all of those markdown core files that we just discussed like the ages.md, soul, tools, identity, user, heartbeat, and even memory. What actually happens with OpenClaw is that every message that you send it, it reinjects those markdown files as a prompt along with the message that you send before responding to you. And this is a big part of the reason why if you're doing a pay as you go model and using an API key to power your open claw, it can get very costly really fast. Especially if all of these markdown files are, let's say, 10,000 tokens in full. Then every message that you send, you automatically incur 10,000 tokens because all of these markdown text files are always injected every turn. And is a core reason why generally I would recommend doing an ooth configuration so that your costs are predictable and capped every month.
Now, a related concept for you to better manage your costs would be open claw's ability to accommodate a lot of different AI models out there. It's model agnostic, so you can think of it as a cockpit that can switch engines mid-flight. Basically, a lot of people run it with Claude from Entropic, but you can even use OpenAI's GPD models and even fully local models through O Lama that cost nothing to run, assuming you have the right hardware to run them. And remember that different models reason differently. Claude's models might handle a complex coding task better, but maybe local open source models, which can be cheaper, can be assigned for simple tasks. And so what you can do is to pick the default model that fits the task of your specific AI agent. But but what's key is that with OpenClaw, you're never locked into one provider. Now, if you're curious what the default model is for each of your agents, you can also view that in the gateway web interface under here in the overview tab.
Now, let's talk about skills. Because out of the box, OpenClaw is a general list. It can do a lot of things, okay? But for specific tasks, particularly for tasks that may be unique for your own use case, it can be exceptional if you give it a playbook. And skills are exactly that. These are basically pre-written playbooks that teach the agent how to do something specific. And each one is defined by a skill.md file that is just another text file that includes prompts, resources, and references. Now, each of your agents can have their own set of skills that they can have access to. And for you to find that, all you need to do is to go to the skills tab in your gateway. And here you'll be able to find the skills that it has access to. You can also go to this specific skills tab at the sidebar, and that will just give you a view for all of the skills across your workspace. Now, if you click on this browse skill store, that just links you to clawhub.ai, which is where people upload and publish their own skills that others can use. Now, because these are community-built skills, that means you need to be really careful about what you install from here because not every skill has been vetted. So, if you're prioritizing for safety and especially if you're a beginner, I suggest that you veer away from Clhub at the moment and either build your own skills or get them from more trusted sources.
Now one thing you'll notice is that you can't actually browse the full text of every skill file directly from your gateway. So, what I built for our OpenClow workspace to give you an example is this skill tree view. And essentially, it's a single graph chart that shows all of my agents in here. And when I hover over them, I can see which particular skills they're connected to, their description, as well as what model they're running. And if I click on a specific skill, I get the description that gets loaded into the agents context. And I can also view the actual skill markdown text to see exactly what it contains. And so, this view is really useful in seeing how many skills your agents have and spotting which ones may be clogging up your agents context. And by the way, if you want this view for your own setup, I included the full prompt to set it up just in the link in the description below.
Now, the next concept are MCP servers. And this is important because your business probably doesn't live on just your laptop alone, right? It's spread across Google Workspace, Notion, maybe you're using Figma and other tools. Now, MCP stands for model context protocol, and they're basically like universal power adapters for your agent. So, each one plugs your OpenClow agent into a different external service. If you connect an MCP server for, for example, Google Calendar, then the agent can read and create events for you. If you connect one for GitHub, then it can manage your repositories. The agent basically reaches into all of those tools from your conversation. One example of an MCP is this one by Blender where in this video Riley basically connected his Open Claw to the Blender MCP so that it can use that software by itself.
So, we talked about skills, which are basically playbooks or instructions written in plain English that teaches your agent how to do specific tasks. Now, plugins are one level deeper because these are actually code level extensions written not in plain English. Usually, they're in Typescript or JavaScript. So, actual code that hook or plug in directly into the gateways internals. So, basically within OpenClaus code design and database itself. And because they run inside the gateway process itself, this means that they can do things that skills simply can't. So plugins can add entirely new messaging channels, for example, register custom tools and even swap out the context engine and the way it manages context that we talked about earlier. And to make this easier to understand, every channel that you connect like Telegram, WhatsApp, Discord, those are actually plugins under the hood. When someone builds support for a new platform, what they're essentially doing is writing a plug-in. If you need an example of a plug-in, there's this recent one called Lossless Claw that aims to improve context management for your agents. But as you can see here, this is written in Typescript and Go. And since this deals with your open clause code directly, then this is a bit more of an advanced feature that you should check with your agents first if it makes sense for you before you install it. But if you were to install it, the process is the same where you just give the link of this repository to your agent and it will figure out how to set it up.
Now, a concept that we will probably see more in the coming months to years as AI agents develop further is this idea of nodes. Because right now, most likely your agent is living on just one machine, right? But what if you wanted to reach into your other devices? Well, nodes are basically those devices that you can have your Open Claw connect to. So, if you set up a paired node, for example, to your smart glasses, then theoretically, Open Claw can see what you see. Or if you set up your iPad as a node, then it will be able to do things in the future like push a notification straight to your iPad instead of sending it through your messaging channels like Telegram or WhatsApp. Now, obviously, since this tech is so new, there are very few examples of nodes being used at the moment. And right now there are only some experiments like this one called vision claw that has that use case of connecting your open claw to your smart glasses. But I'm pretty sure that as this technology develops we'll see more of this implementation really soon.
Now a very important topic to take note of is around permissions and security because yes open claw is powerful since it has its own computer and you can even connect it to the tools that you use in your business that are living online. But the flip side of that power also means that the security stakes are quite high. And so a very simple tip that you can do here is to just ask your agent to set up a cron job that will do a security audit for your own specific setup. But also remember that you can actually control what the agent can touch in your workspace files. There will be a file called openclaw.json which you can just ask your agent to edit for you or you can edit it yourself to explicitly allow or deny specific tools. So if you don't want your agent browsing the web, for example, then you can just deny the browser feature.
And there you go, 20 plus Open concepts explained in plain English. And if you want to learn more about Aentic AI, then watch this video next. But that's it for this one. I'll see you guys next time. Thank you.
[music]