Transcription
OpenC is one of the most important and powerful AI tools of all time. There's no way around it anymore. You have to engage with it. I myself have dealt with it extensively. I have spent well over 100 hours with OpenC, have learned it inside and out, and in this video, I will give you all the most important details in a compact guide. Whether you are a complete beginner or already a bit more advanced, you will learn something new in this video. Promise. Enjoy. It's quite crazy, folks, how quickly all this has happened. Started as an inconspicuous open-source project, suddenly went viral around the globe. It has now become the number 1 software project on GitHub. It has more GitHub stars than React or Linux, and that in such a short time. The AI industry has completely shaken it up. Open AI has, as is well known, brought the developer of OpenC into the fold, and Nvidia has just launched Nemo Core, and this is all just the beginning. OpenC is not a new AI, let's get that out of the way. It's not an AI in the sense that GPT or Opus or Kimi is an AI. OpenC is much more of a bridge between AI and the rest of the world, all other tools. It uses a language model like Opus 4.6 as its brain. It has an external memory in the form of Markdown files. It has a heartbeat that allows it to act proactively, more on that shortly. And above all, it has access to potentially all your tools and even your computer. Yes. That's always the big debate, how and where to best install OpenC. There are three options. First, on your laptop. Yes, you can easily install OpenC as a local program on your own computer. It has the big advantage that it is completely free. It's an open-source tool that you install on a device you already own. Personally, I still thought it wasn't a good idea, not because of security concerns or anything like that, about accessing your own files, but because it's simply a bit impractical. My agent should act independently and be accessible at all times. It's more practical, therefore, to install OpenC on a computer that is always online and statically accessible via the internet. In other words, on a server. That's the second option, a Linux server with OpenC in a Docker container. I would recommend that to IT professionals, and I would even go so far as to say that every IT professional needs a server. But for non-IT professionals, there is now an interesting third option, and that is Managed OpenC. The same principle as, for example, with Managed WordPress, where you don't get access to a server with all its possibilities and all the responsibility of administering it, but OpenC is managed for you by the server provider, and you never have to worry about updates yourself. I don't use it myself, I'll be honest, because I'm an IT professional. Nevertheless, it's certainly interesting for many of you, and both options 2 and 3 are available at Hostinger. They support this series of technical videos on my channel, and it's also a good way for you to support my channel. I'll put the link to this page in the description, and then it says, IT professionals, please turn right, not IT professionals, turn left. Either self-managed OpenC will mean a Linux server, or managed OpenC will probably also mean a Linux server, but one you don't have to worry about. The whole thing is still in beta, but it seems to be working already, and they are constantly adding new features, for example, to make setup easier, so you don't have to look into this complicated OpenC gateway anymore. I would say, decide for yourself how IT-savvy or nerdy you feel. The next decision is then, which AI do we use for OpenC? The answer. Honestly, many paths lead to Rome, and Hostinger also offers the option of having the AI set up immediately. So you can now have the AI tokens billed directly through Hostinger and then even exchange the AI model as you wish, without ever having to connect external providers. That's practical and expensive, I must say for the sake of transparency. They use Nexos AI, which then provides the models from Anthropic, OpenAI, and the like, and naturally takes a certain markup as an intermediary. It depends a bit on your situation. If you have no financial worries at all and just want everything to work out of the box, and preferably with the best possible AI model, then this is exactly the right way, and it's best to choose Opus 4.6 as the model, as it's the strongest. If, on the other hand, you are willing to pay for tokens but would prefer it to be a bit cheaper, then honestly, it's better to create a separate account with Anthropic, connect the API key, and then perhaps not use Opus, but for example, Son 4.6. It's not that much worse for most OpenC global use cases and significantly cheaper. If you're a real bargain hunter, then I would say, definitely use an AI subscription and not an API key where you pay per token. So, if you already have a ChatGPT subscription, for example, you can simply connect it to OpenC. Cloud subscriptions might work, but are officially forbidden, and some accounts get blocked. I don't know, dude. However, the Olama cloud models work great with OpenC. I've made a whole separate video about it on the VPS. My secret tip, if you're not entirely sure how to connect your ChatGPT subscription to OpenC, then install OpenC first with these built-in Hostinger Nexus credits, and then, as the very first thing, tell OpenC, almost as a greeting: "Hey, OpenC, please help me connect my ChatGPT subscription." It's a bit like using Internet Explorer to install Chrome. In that case, the next step, after the AI is running, is to really get to know this new AI employee and train them, just as you would with a real employee. Take the time for it. I can really only recommend that. Don't expect OpenC to solve all your problems in the first hour, but take your time. OpenC is a tool where the greatest added value lies in working with it long-term, getting to know the agent better and better, and giving it the opportunity to get to know you. Who are you privately, professionally? What are your goals? What are your biggest bottlenecks right now? What are your communication preferences? How do you imagine your OpenC could help you, and what are ways it could help you that you haven't yet imagined? Yes, it's worth asking that. This is called reverse prompting. Before OpenC, with ChatGPT and similar tools, there was always talk of normal prompting. How do I formulate my request to the AI particularly well? Now, the tables are turned. Instead of saying, "Help me with this and that," you ask the agent, "How can you help me?" Based on everything you already know about me, what would be good ways for you to generate added value in my life and for my business? Or if you don't know enough about me yet to judge, then ask me questions again. In the course of such dialogues with your agent, OpenC essentially creates a profile. It remembers your situation, your preferences, it gets smarter with every conversation. How does it do that? Amazingly simply with text files, with Markdown files that are simply located in the OpenC directory. And that's the next tip, I recommend that you take a look at them yourself. These are the most important files that make your OpenC what it truly is. Let's take a look. If you're an IT professional and run OpenC on a VPS, you might even have the files mounted in your local file explorer. Pay attention, folks, I have to demonstrate this once. Here Finder, this is my computer, this is my Hostinger VPS, and you can see into the OpenC workspace and then have all the files like Soul.md or Tools.md alternative. We go into the OpenC Gateway, then here once on Agents and Files. I see right now, I need to update my OpenC. Important tip: regularly update your OpenC. Simply click on the three dots in the Age panel and update. So the UI has changed a bit again, but the content is the same. Let's go into Soul.m, for example. Here the soul of your agent is defined. "You are not a chatbot. You are becoming someone." What does your OpenC's soul file say? Does it precisely describe the assistant you want, or are there additions? In my case, for example, whoever saw my last video will recognize this here. "No empty apologies or promises, please." These Markdown files can change over time by giving your bot feedback. Or you can also manually change them. It ultimately makes no difference. The bot won't suddenly complain in confusion, "Who changed my soul? What happened?" No, there isn't one bot. There isn't one agent. In the last video on my channel, I compared it to the Christopher Nolan film Memento. Every day, a new Memento OpenC agent awakens, basically without any memory, and reconstructs what happened based on its Markdown notes. That's why it's all the more important that these notes are accurate. Well, Soul.md won't change constantly, just like User.md here, although it thinks I'm still in Istanbul. Maybe I should let it know, I'm now in Hamburg, and I see that my Hostinger server is in Kuala Lumpur. I set that up when I was in Thailand, but in terms of ping, it doesn't make much sense anymore. It's nice, folks, I always find something new out when I record a video. Can that be changed? Please tell me it can be changed. Settings. Ah yes, here Server location Kuala Lumpur. That could be changed to beautiful Germany, for example. Ah, but that would be a completely new server, wouldn't it? Without my data. Okay, I don't have to pay again. I'll probably do that at some point, but not now. But that's also the beauty of OpenC, by the way. I can download these Markdown files, the memory, and load them into a new OpenC agent. They are just text files, and then the new OpenC agent knows as much as my old one did, or rather, it even thinks it's my old agent. Is it then my old agent, or just an identical copy with the same memory? Is there a difference? Philosophical questions. Okay, back to the topic. Files like Soul.md or User.md shouldn't change constantly anyway. In contrast, Tools.md or Memory.md are more dynamic. The Tools file contains important information about special tools and services that your OpenC should interact with. For example, if we look at the Tools file of my main agent, Amadeus, then we have here, for example, the information that it should use SuperData to transcribe YouTube videos, or here the IP addresses of all Hostinger servers, because I really don't just talk about Hostinger in any product placements like some creators, but I've been hosting everything at Hostinger for a long time. In Tools.md, you'll find where this or that API key is located, or which parameters it should use for its Reddit research. All such technical things go into Tools.md. All other general information goes more into Memory.md, for example, the rules for your collaboration, or the details of important key projects, or your quarterly goals, all such things that are worth keeping in mind for your OpenC in every session. Otherwise, there are also daily notes, the OpenC diary, where it generously documents everything it does and calls itself throughout the day. In the last video on my channel, we went into a bit more detail on that, and I can really only encourage you, take a look yourself, get hands-on with these files, especially if you've been using OpenC for a while and have never looked into it, it will be very interesting for you to see what it has remembered, and always keep in mind during interaction that it can write in there. You can even explicitly trigger it by writing to your agent, "Remember this for next time." It really helps to understand how OpenC works under the hood. And the last file I want to point out to you is Agents.md. This has now become a convention. Many tools have such a Markdown file in which the LM is explained what it has to do in the first place. This is also not usually edited. So your Agents.md probably looks the same as my Agents.md. You can see here how OpenC onboarding works, which files it looks at first in each session, and here really just a lot of information on how OpenC proceeds exactly, and especially down here, also some information about heartbeats and cron jobs. Homework for you: read through your OpenC's Agents.md thoroughly. But yes, I absolutely wanted to talk about this one topic. The heartbeats and cron jobs, along with the Markdown files, are OpenC's absolute unique selling proposition. They are probably the reason why working with OpenC feels different from working with many other AI tools. So, Cron itself is first and foremost an ancient scheduling tool for Unix systems. In principle, it's simply about recurring appointments. If you tell your agent, "Please send me a briefing on topic X every morning," then it will acknowledge a cron job. At a specific time, OpenC's alarm clock rings, so to speak, and then it follows the instructions stored in the appointment. We can also look at that in my dashboard. If we go to cron jobs here, daily open glory search, once a day, follow the instructions in this file. Let's take a look at the file. Where should that be? Here Workspace, and then Tools, and Daily Research Workflow. There it is, as a Markdown file, a step-by-step guide for the OpenC research flow, which leads to OpenC looking at, for example, the tweets of Peter Steinberger and the like, and sending me a summary every day. Cron, therefore, is the mechanism for everything that recurs and should happen at a specific time, once a day, once a week, and so on. Even more interesting, however, are probably the heartbeats, which are the mechanism that truly brings OpenC to life. A pure ChatGPT always only reacts, but OpenC also becomes independently active. So, it's often said, OpenC is like a 24/7 AI employee. That's almost correct. More precisely, 24/7 every 30 minutes. Yes, every 30 minutes, a heartbeat occurs in OpenC, which can also be adjusted, by the way. And with every heartbeat, OpenC looks at its Heartbeat.md file and then simply follows the instructions. If we take the Heartbeat.md of Amadeus here, for example, then we see that there is a synchronization script that it executes. It checks if its diary is up to date, and it also performs an uptime check of my three most important websites, for example. Yes, the heartbeat is what enables me to say, "OpenC, monitor these websites, for example," or "As soon as Tagesschau posts an article on this or that topic, let me know," or "Do the following." It won't do it in real-time, but with a delay of up to 30 minutes. But yes, it's precisely the heartbeats that transform a completely normal reactive LM into a proactive, independent employee. It should slowly start to click for you how you can use this for your personal use cases. I've made a whole separate video about OpenC use cases. From programming to research to WordPress maintenance, I've explored several options in detail. I'll link it again afterwards. For some of these use cases, it's of course important to give OpenC additional tools. We already saw that with the daily research, where it used a Reddit skill. Skills are essentially extensions for OpenC. However, here comes my hot take. Are you ready? Skills are completely overrated. It always sounds so cool. "I've installed a new skill for my AI." It's like some ultra-cool plugin or something. You know what? Pay attention, let's look at these great plugins. Chwhub. What do we have there? Here, for example, from Steinberger himself, a Trello skill. Wow. 1000 people have installed it. VirusTotal also says it's benign in OpenC's own algorithm, it even flagged it as suspicious. Not that there's any malicious code hidden in the skill, but wait a minute, what is the skill? Skill.md, a text file, and do you know how long it is? It goes from here to here. That's the whole thing, not the summary or the description. The entire skill, printed out, fits on one page, and I'd argue that if I simply tell my OpenC, "Please use Trello," it will manage. If necessary, it will just quickly search the internet for the official API docs. The skill is ultimately just another memory file, and thus indeed a useful thing, so that it doesn't have to search the internet every time again, after it has, in its forgetfulness, forgotten everything from the previous session. Memento, remember. In that case, it's quite good that it documents such APIs for itself, and instead of putting everything into Tools.md, you create individual Markdown files per tool, and they are called skills. Useful, absolutely, only please don't let yourself be misled by this cool word, because I have the impression that some people think skills are more than just that one short text instruction. I honestly have some security concerns, but in case of doubt, of course, don't download any instructions for your agents from the internet, but let your agent create the skill itself. If you have worked on something with OpenC, and it has started using an API successfully, then tell it briefly, "Please remember this for next time as a skill," so that it doesn't have to research the internet every time. I'll get to the topic of security shortly anyway, but wait a minute, I'm realizing I keep saying your agent looks something up on the internet, it uses some API, but how does it do that at all? That's a very central feature, that your AI has access to the internet, I should probably mention that in this guide. The standard way is the Brave Search API, simply because it's free. At least it used to be. They've probably also made it a bit worse due to the increasing popularity of OpenC, I believe. Until recently, you could simply get a free account and had, I think, 2000 search queries per month for free. How does it look now? Brave Pricing. Ah yes, now we have $ per 1000 requests, but also $ free every month. Why they formulate it so godlessly complicated is beside the point. But I'd argue that if you set a $ limit in your Brave account, it will effectively remain free. So that's still a good option. 1000 searches per month isn't that little either. Otherwise, another option if you're willing to pay at least a tiny bit for internet search is Perplexity. Some of you probably already know it, an AI specialized in search, and yes, you can also use it via API. The motto is always the same: create an account, link payment details, copy the API key, and then simply give it to your OpenC agent and tell it to set itself up. Just be mindful of your costs, as always. If you want to be absolutely sure and have a cool, hip bank account with a good mobile app, then use virtual cards, for example, set appropriate limits on your bank's side, that's a good option. Otherwise, you should also always have the option to set limits directly on the AI platforms so that your diligent AI employee doesn't make you poor. And at this point, another hot take. The biggest danger with OpenC is you yourself. OpenC has an enormous layer of vulnerability, as we IT professionals would say. Yes, it's an attractive jackpot for hackers because, just like with N8N, for example, a lot of credentials, passwords, API keys are stored in one place. Yes, especially if you run OpenC on a virtual server, especially if you might be using a custom setup with Olama Lama or something, as I've demonstrated, cool and cheap, but such open-source models are naturally more susceptible to prompt injection. Yes, in such cases, you should definitely familiarize yourself with OpenC security. The good news is, you'll find a very comprehensive video on my YouTube channel where it's all about OpenC and security. And also one on N8N security. I take the topic quite seriously, so feel free to look around. But in my opinion, the biggest danger for most end-users experimenting with tools like OpenC is rather doing something wrong yourself, giving wrong instructions, so that in the end the OpenC agent suddenly deletes the entire email inbox, as happened to one of my colleagues, or that it sends out a completely embarrassing email in your name. In my opinion, especially for those using managed OpenC with a modern LM, this is the more realistic danger than being hacked externally. What protects you in both cases are primarily common-sense principles. Don't give OpenC more rights than it needs for its tasks. Treat it generally like a potentially fallible, but new employee whom you don't know well yet, and to whom you probably wouldn't give your private credit card. If you behave reasonably and cautiously, then yes, you will certainly enjoy working with OpenC. And if something does go wrong and OpenC no longer does what it's supposed to do, or it stops responding altogether, then here's another important tip: simply turn it off and on again. Seriously, in most cases, you don't even need to restart it completely. That can be done in the Age panel via the three dots and then restart. In most cases, however, a simple slash is enough. I say it like that, it's enough, but I admit, folks, my OpenC recently completely froze, and after trying a lot, trying for ages, and looking in the server logs to see what might be going on, I remembered again. Snow, just make a new session. Many people don't even know about these slash commands. I can type a slash here in Telegram, and all sorts of useful OpenC commands come up. Slash stop, for example, if an OpenC is currently doing something it shouldn't be doing, you can interrupt it with that. Otherwise, I can only say about troubleshooting, if something isn't working with your AI. In the OpenC Gateway, you'll find logs and even a whole debug tab where a lot of technical information is provided. Even if you don't understand it, you can paste it into ChatGPT and ask for help. Or some people even do this, they install something like Cloud Code on their VPS and then say, "Hey Cloud, here's a broken OpenC, please fix it." Or, in my opinion, the best option: just get another OpenC. You can simply open a second instance. I could go into my Docker catalog here, OpenC, and just install it again. Then I have two independent OpenC instances side-by-side on the same server. I'll install OpenC on my own computer temporarily, just for once, to then tell this fresh OpenC about the problems with the other OpenC. It knows itself best, so that has always worked best for me. What do you plan to do with OpenC? How will you use it? Here you'll find a video with my OpenC use cases, and here the video on OpenC security. If you have further questions or ideas for cool videos, please let me know. Sometimes I make it a bit more technical, sometimes more beginner-friendly. I try to vary it a bit. In any case, I'm really enjoying making the videos right now. I hope you are too. Take care. Until next time.