This box is set up as a dedicated 24/7 worker and we're going to set one up for you as well. Think of something that is done over and over in your business or personal life. A repetitive task, maybe, you know, checking emails or drafting invoices. Now, try and explain why you do the task the way that you do it. What do you look at? What do you ignore? When should you flag something to be checked by somebody else or somebody higher up? If you can explain that thought process, chances are you might be able to start automating a chunk of it with what we're about to do.

This is our NVIDIA DGX Spark and it's running a local AI agent with OpenClaw.

Hello, Jared from the future here. Now, in your time, OpenClaw is probably not, you know, the latest and greatest framework and we're not here to hype you up about OpenClaw. We're just more so doing AI agents in general, but the skills that we're learning here are transferable to Hermes and probably most AI frameworks. We're just trying to get you up to speed on AI agents at a whole.

This agent runs offline. No subscription, no data leaving your building. And it makes you less reliant on big AI company data centers. Well, we are going to be installing Claude Code to, you know, set this all up, but once it's running, the day-to-day tasks and automations are all done locally. The models that do the thinking are open weights, sometimes even open source. They think on the machine and they're yours to keep.

Now, in a previous video, we set up Ollama and Web UI on this, which gives us a nice little endpoint of essentially running an LLM offline with a nice, you know, web page interface you might be used to. If you haven't seen that, it might be a smart idea to go and watch that because we are going to be taking that chatbot LLM experience [music] and using OpenClaw to turn it into something that sort of exists.

OpenClaw is the real star of the show here. It's a fancy piece of software that is going to let your AI remember things like how you like to solve problems, the way that you want it to work, how it should speak and act. And it also, you know, lets it remember important facts about us and our environment and our business and just things around us. It will also give it access to tools that allow it to read and create files, access and search the web, connect to your email or, you know, Google services, read and create spreadsheets, and with this it can monitor fuel prices. It can draft invoices. It can write code. It can summarize scribble down notes or maybe a meeting. It can track stock and inventory. It can handle support tickets. It can organize task lists. It can monitor real-time news on critical topics and so on.

If this is your first time diving into AI agents, it is going to feel pretty darn sci-fi. And if you're like me, you're probably going to be, you know, annoying your friends over the next few weeks by telling them all the fancy ways that you've automated the mundane and really repetitive parts of your life. So, let's get into it.

Starting off, let's open a new terminal on our DGX Spark. Remember, we're using NVIDIA Sync here. Really handy app. Now, there is a nice Docker container that you can pull and go and set this up manually, but we live in 2026. Well, I do currently. I don't know what year you're from. And the best way to get an AI agent going is with another AI agent, but something more powerful. We are going to be using Claude Code for this because it's you know well it's what we got and what we use but Codex could do this just as easily or you know any other frontier powered coding agent that exists in the future. Just be sensible and use something reputable to install Claude Code. We're just going to go ahead and curl it with this command here. And then we can just go ahead and punch in Claude to start it. All of the commands and prompts that we'll be using in this video you'll be able to find in the full written guide linked below. So head over there if you need it.

Now we are just kind of vibe jamming some stuff out here. So, I'm going to kind of do it and then recap what we did and paste it all together so you don't have to watch, you know, 40 minutes of terminal lines go by. So, we're going to go ahead and run through our first time setup and then we're going to, you know, very importantly choose a theme that we want to be using for our terminal. And then, you know, we'll go ahead and set up all of our account details and sign in. This should be pretty straightforward for whatever coding agent you're using. And you might need to go ahead here and approve letting it access to your files on your computer. Obviously, you know, if you really want to be safe, don't have anything on your DGX Spark that's, you know, critical or sensitive right now.

Now, everything we're going to be doing here is relevant to Claude Code, but it should be transferable across other agents. The first thing we want to do is select our model with a /model. We're going to be using Opus here on high effort. We don't need anything too crazy powerful here, but we do want something that is capable. Then, we're going to go ahead and paste in our prompt that you can find again in the written guide. That's the last time I'm going to say it. And let it run. Now, this prompt is giving context to Claude about what we did last video and how it [clears throat] should set things up. This just, you know, makes it easy. But you could just have easily have, you know, told Claude to poke around and see what it can find and get it to install and plug in OpenClaw to our Ollama instance without this prompt. That's one of the the beauties of this stuff.

And after a bit of work, Claude has come back telling us that Gemma is running a tad slow. This is sort of normal and it has gone to the liberty of benchmarking a similar capability Qwen model that we had installed. This is a bit of a spoiler for the end of the video and a future video. By the way, once it's gone through and set up what it can, it's left us with some human tasks here, like accepting a security disclaimer and going through the first time setup. So, let's go ahead and punch in the commands that Claude has told us to do. Obviously, you know, maybe worth a double check. Not that Claude's doing anything nefarious, but always worth a double check. Be a bit security-minded there.

And we're going to go through and run our open claw first time setup. If you are unsure about any of these, you can ask Claude. That's why it's there. But for our needs, we're going to go ahead and run the quick start setup process. We're going to go ahead and skip the model provider. And we're happy to set our default model to, you know, the Gemma one that has already been set up.

Now, we can go ahead and set up a channel. A channel is just a messaging service or platform that your bot can interact with you on. We're going to be using Telegram, which is a really common one, but you can choose Slack, Discord, or whatever is most convenient and integrated with you. Pick the one that is, you know, the most accessible for you. Go ahead and set up a bot on your platform of choice. We've just gone ahead and set one up for Telegram, and we're just going to punch in that bot token like so.

Now, we can set up a search engine provider. Right now, this is not needed, so we will go ahead and skip that. Our bot will still be able to access the web, but you'll need to give it a direct readable link. Essentially, the services that you see it there give it web searching capabilities to actually find a link. Your bot can't exactly do a Google search natively like we humans do. You could feed it a direct link and it can try and read it, but it can't, you know, get those links. Not all of these search providers are free, though, so watch out. But SearXNG is a really good locally hosted option that Claude can set up. Maybe chase that down after this video.

There's a repeating pattern here in case you can't tell. Open Claw is going to be our agent that runs all of our tasks and automations and Claude will be organizing and managing the infrastructure that it uses. You could also just forgo Claw and, you know, use your brain or something like that and try and set up that, but I don't have time for that. It's 2026.

Here, we also have the option to install some extra skills that come with our agent. Choose these if you want. We're just going to leave them for now. We don't really need any of these. It's also going to prompt us to set up some keys that we don't really need here. Again, if you need them, you can set that up now or you can set it up later. And finally, we are going to enable two hooks here that might come in handy. Command logger and session memory. These will just make our OpenClaw agent run a bit better. And you know, it's a bit nicer to use. And finally, finally, we'll go ahead and hatch our agent. And we will do so in this terminal here. And with that, OpenClaw is up and running.

Now, right now, this very instance, we have a brain capable of some pretty serious thinking, but it's a complete blank slate. It is literally this thing's first day on Earth. So, we need to start teaching it who it is, who we are, and what we are trying to do. How do we do that? We just talk to it. Just talk to it like you do a regular person. I'm going to go ahead and tell it what it's running on, as well as, you know, some rules that we'll give it about how it should do things. And I'll also go ahead and get it to give me some questions if it needs anything else. Now, what you do here is up to you. You don't have to follow what we're doing. This is your chance to, you know, set up your agent and mold it to behave and think the way that you want it to do so. So, maybe spend some time here or go ahead and find some really good meta prompts for this setup.

I'm going to do something here though and ask it to run me through the files that it runs on. And this is going to reveal a little bit of the magic that happens behind the curtain. So, OpenClaw gives our LLM access to tools like web searching, accessing files, and you know, our Google account, that stuff we talked about before, but at the end of the day, a lot of this is thanks to simple text based markdown files that it just reads. For example, it stores its personality and how we want it to work in its soul and identity files. In the user file, it's building kind of a database of, you know, facts about us and the world around it and who we are and what we do and whatnot. The memory file is just a long running file with things that it's, you know, deemed important to remember. And our agent is really just an LLM with this data injected as a part of its core being.

If we want to change how it behaves, we just tell it to do so. And it will update these files by itself for future reference. That's really kind of all it is. It's a whole bunch of tools with files and an LLM. and then the LLM reads the files and then you talk to the LLM and then it kind of updates the files for the future and it gets better over time. Want to teach it to do something? Give it a prompt or you know run it through the steps or the thought processes of a task and get it to save it as a skill that it can search for and read later. The real trick there is that is usually smart enough to write these down itself.

Now you can use the terminal if you already want to look like an elite hacker, but let's get something better. There are more handy ways to use OpenClaw. First of all, let's get Telegram working. Previously, I went ahead and set up the bot for my agent. I'm just going to go to Telegram and hit it with slash start. It's going to tell me that, you know, it's currently not set up. And then I just need to punch this pairing code into my OpenClaw chat like so. Now, I should be able to just go ahead to Telegram and message it. Hey, very cool. Again, this will depend on your channel that you're using. Discord might have a different process, but it should be pretty straightforward.

This also does bring up something important and that is chat instances. The instance of Sparky that I'm messaging over Telegram is different to the one in the terminal. Our instance in the terminal is not currently aware of the message that we just sent via Telegram. That's because they're both sort of clones of our agent if that makes sense. They share the same soul, the same memory, the same identity, the same, you know, skills. Everything about them they share but they are different chatting instances spooled up from those memories and those files. If one of them goes ahead and writes a skill then the other one can you know see that and get that but chat-wise they're largely unaware of each other until something is written to file.

Now we are interested in further developing our agent from the OpenClaw dashboard. If we were directly on the DGX Spark this is an easy task as we can just get you know a URL to punch in but that URL is local. It only works on the DGX Spark and I'm on another device. So, we're going to need to create something called an SSH tunnel. If you know how to do that, go ahead and do that. But I'm just going to go ahead and explain to Claude what's going on here and just ask it to give us instructions of how to access that. It's gone ahead and given me some commands to punch in. I'm going to do that. That's just standard SSH stuff. And I'm going to punch in our web browser URL and come to look at that our dashboard.

This is a nice and fancy bit of UI. And as you can see here, we can see our terminal chat that we just had before as well as our Telegram instance. And we can keep chatting from, you know, here in each of these instances if we want. Along the side here is a whole bunch of helpful management tools to check out as well. Now, how you choose to interact with OpenClaw is up to you, but the dashboard might be the best way, but I do like to have Telegram or some sort of a messaging service to build things from my phone when I'm out and about. The other reason the dashboard is handy is that we can see all the tool and skill calls. And basically anytime it you know writes to file accesses a web page and this is super helpful as we can see what it's trying to do when we give it something to do.

While we're here let's talk about tools and skills. Tools and skills are some of the foundations of OpenClaw as well as other agent frameworks like Hermes. Tools are basically just capabilities that you equip your agent with. Accessing a website is a tool. Reading and writing files is a tool. Accessing a specific API is a tool. Skills, on the other hand, are a set of instructions on how to do something. Using these tools to accomplish something, how to draft an email in our company style, a skill, how to try out your support ticket, a skill, how to ask the user a series of questions to scope out a project according to some framework and break it down into manageable tasks, a skill as well. And all a skill is is just a markdown file of text with instructions that you know your AI reads and goes, "Ah, that's how I do things."

You can teach and train OpenClaw how to do things and get it to store it as a skill or you can search around for skills online that people have already written. There are skill repos like ClawHub to check out. You can find them on GitHub and usually you can just paste in the link to your OpenClaw agent and say learn this skill or maybe like hey I like this part of the skill. Can you adapt this existing skill to it? Just be aware though you are giving your agent instructions to execute from a stranger on the internet. So look for trustworthy sources. verify. You know, there's a little ClawHub security audit. You know, double-check on that. Use your brain here. You know, time in the sun is really your friend in this field. Be responsible and be security-minded here.

Speaking of, let's go ahead and create a skill and also set up a nice little automation while we are here. Now, every single person watching this video is probably going to have a completely different use case for an AI agent. So, we're going to do maybe something that's a bit of a, you know, universal demo, a morning news brief tailored to your needs. So, we're just going to keep talking in our main instance here. And what we're going to do is we're just going to let it know that, hey, I'm going to be prompting you in two steps here. Then, I'll go ahead and paste in the second half of the prompt, which is just kind of a like a bit of a meta prompt that, you know, tells OpenClaw to run us through and set this up, you know, with us and our, you know, ask for our preferences and whatnot. Shout out to NVIDIA. We are kind of borrowing this prompt from the open claw guide. More on that later though. And here I just ran through all the questions that it was asking. You can be as detailed as you want. You can add exceptions in here. This is your chance to offload, you know, the exactness of a task that you want. These are easy to update later as well if you just want to keep it light for testing purposes. And as instructed, it's gone ahead and spoiled up a demo morning brief. Pretty cool.

Now, that seemed like it worked, right? Well, it didn't. It got close, but there's a few things that went wrong there. I didn't see any, you know, files being written. The skill was not saved here. We could go ahead and manually ask it to, you know, save that, you know, thought processor skill and manually check that. But some other things have gone wrong here. For example, I didn't see any tool calls checking those links that we gave it, nor did I see any cron jobs being created. Cron jobs are just a task that is fired up at a given time. For example, our morning news brief should have had cron jobs created for the times that we requested the morning news app. Also, this news is hallucinated. Everything that's happened here is either made up or it happened years ago and you know, the model's been trained on it. It's just regurgitating it. What are the odds of the Western Australian potato farming industry and Mongolia both having major news stories today? I threw in these news interests as just a little bit of like a red herring here for our model.

And this raises an important point. Test your systems. especially AI ones and especially ones that run, you know, offline models that might be more, you know, prone to hallucinations. And if you're developing, you know, from the top down, you've got all these fantastic AI tools that let you roll out all these things really quickly. It's very easy for it to appear that it's working. So, test test.

So, how do we fix this? Well, we can just ask it in plain English. I'm just going to ask it to go ahead and manually check those new sources for us. And as you can see, it successfully used the web tool and we successfully got some front page news. Interesting though, after it checks, it gave us a generic kind of welcome message there. I think we have a bit of a context window issue here or something went wrong with our model. Now, the model and our context window things are part of our kind of AI infrastructure stack. So, it makes sense to switch over to Claude and get it, you know, to do some digging.

Claude has access to OpenClaw as well as Ollama and the gateway between those two. But because it, you know, has access to, you know, the start to the end of this process, it is really good at helping in situations like these. After a bit of digging, Claude eventually found the culprit. When we set up Gemma and Ollama last video, it only gave it a 32k token context window, which is kind of I think it's an Ollama default, but OpenClaw thinks that it has nearly a 200k token window. It's given us a few solutions here. I'm just going to go ahead and ask it to, you know, max out Gemma's token window. We've got the hardware to do so. You know, we might as well just do that and fix the mismatch.

If you used a more powerful coding agent, it might have detected this mismatch when we went ahead and set this all up initially. Or maybe you installed Ollama and the model with Claude and not manually, and it did the right thing the first time. What I'm trying to say is you might not be getting the exact issues that we have here at all. We kind of just planned this for educational purposes. Then we're going to go through and you know update the initial prompt to you know flag this issue and so that you don't run into this. But this is just a great little example of you know having your coding agent working alongside you know your open claw to help set up things and you know diagnose and whatnot. If you really want this issue you can probably go ahead and remove that line from the initial prompt.

That also leads us to some important questions. What size context window do you need for an agent? Well this depends on what you're doing. You know how long is a piece of string? It also depends on how complex it is. I hear some people say you can get away with 32k tokens. It is possible to run that small, but usually 64 to 128k is a bit better of a number. Larger sometimes might hurt as maxing out a context window can kind of degrade retrieval accuracy, but it depends on the model. We are lucky that we have the hardware to usually max out the context window of a given model, so we might as well, you know, at least choose something that is sizable.

This also brings up something that OpenClaw and other agents do called compaction. So we have our context window, we chat to our bot. It loads up all these, you know, memories and skills and fills up the context window over time. What happens when it gets full? Well, it does compaction. Essentially, when the window starts to get full, I think it's about 75% by default. You can change this and check it if you need. It looks at everything in the window, summarizes it, condenses down the important things while offloading older memories and tools and whatnot that it doesn't think it needs anymore. This is a really cool mechanism and it allows, you know, your window to keep being clipped and, you know, an instance can stick around for a long time.

And just some insight into the problem here of what was going on, OpenClaw wasn't compacting the conversation because it thought the context window was huge, but it was tiny. So when it fed new information in all the old information was punched out and it forgot our initial prompt and our sources and everything like that.

So let's go ahead and run that and we're checking sources and we seem to be getting output and oh look at that. All right, it started writing what it can and can't get. And then we get a flat-out error like this. Okay, we are hitting another issue here. I promise we're setting up these issues. Usually it's a bit more smoother, but there's just great learning purposes here. The issue here is that Gemma is just a little bit too slow for the default open claw settings. I kind of suspected this would be the case and [clears throat] it's even baked into the initial open claw setup prompt to you know be like hey this model might be too slow. OpenClaw sets a maximum time for a task to take and usually by default it's about 900 seconds. This is to prevent an LLM from getting stuck, you know, in an infinite loop overnight and, you know, burning 30 million tokens and, you know, we're running a local model, but that's just, you know, jamming up our system and wasting power.

Ah, this sounds like an infrastructure problem. Let's see if Claude can confer. And after reporting it to Claw, it gave us two options. Increase the timeout or swap to a different model. We are going to be playing around with running some, you know, other models in the next video. So increase the timer if you want to keep using Gemma or whatever model you want. But we're going to go ahead and swap to Qwen. Now if you don't already have Qwen like us, you can ask Claude to go ahead and download it. It's got access to Ollama. It can just pull the model for you and just tell you when it's done. And we're using Qwen3.6 35B-A3B Q4_K_XL quantization. I think something like that. Look, just go and find a good Qwen3.5 model. This model is a mixture of experts model. Spoiler alert on an upcoming video. mixture of expert models might benefit the DGX Spark more.

Keep in mind here though, security risk asking Claude to just pull from a random site, you know, just grab a model. You know, there may be a name space conflict, stuff like that. A wise move might be to manually find a trustworthy link for a model with time in the sun and telling it to pull it from there. This is also setting up something cool we'll do in the future video where we set up a model selection tool that can swap the models on the fly. Pretty cool.

Now, let's go ahead and give this another go. To start, I'm going to restart our instance or chat with slash new. Very handy for checking that it can do a skill and not, you know, using existing knowledge that's in its context window. I was also curious at this point to see what it had saved. I don't know what I was doing here, but yeah, anyways, there was something partially there and I think I just said to restart the whole process again.

On this final go, we went through and did the exact same steps again. We put in the sources. It pulled from them. Acknowledged. Sweet. We're getting somewhere. I put in part two, answered the questions for the third time now or something like that. And it looked like it actually worked this time. We've got cron jobs set up. We've got files being written. That is a good sign. Again though, test test. I just went ahead and asked it to fire off one right now. And after a little few minutes, it went ahead and did it. It even reported some issues that it had in that process. It struggled with some links we sent it and suggested some fixes. Ah, now we're starting to get, you know, into AI agent territory. And it's open clawing time. Please don't actually put that in editor.

I'm just going to go ahead and tell it to just, hey, fix it. Hey, that's pretty cool. It went through its own diagnostic and test process, prompted itself to think of some workarounds, found a fix, and then implemented a solution all by itself. I know it's not as impressive compared to what Claude Code has been doing on the background, but this computer was given a task. It found an issue with that task, troubleshot itself, and you know, implemented a fix all with, you know, numerical calculations that were crunching on the machine itself. That's pretty sci-fi.

By the way, if you ever want to change this behavior of the morning news brief, just ask it in plain English. For example, I want it to stop talking about Mongolia and I also want it to, you know, throw next, you know, tomorrow's weather in the morning brief. And that's it. It just goes ahead and rewrites its own files to fix that and carry it out in the future. Want to do onboarding again? Just say so and paste in the prompts again and it will go through and do it and rewrite the files. It's just this is the thing you got to wrap your head around this kind of like new method of development. You just ask it and it will try its best to do so and rewrite its files. Like that's crazy.

Sweet. Let's now go ahead and set something else up for our agent to do. Hopefully, we've got, you know, a little of the bugs and the kinks sorted out and we should be able to just one-shot this one. We're going to go ahead and monitor fuel prices. To do so, we're just going to go ahead and find ourselves a nice fuel API. Now, most states in Australia have their own fuel tracking API, so you might need to do a quick search to find one for your relevant state, but chances are you'll find a free one from your state government, which is what we're doing here. We're just signing up for a free account and a free key.

Now that we have our API key, we need to put it somewhere safe so that you know Open Claw can read it. Now, this is a pretty low stakes key, but in general, should we be pasting a private API key into an internet connected AI? Possibly not the smartest move if we have our security thinking hat on at the time. So, instead, I'm just going to paste in the prompt where we'll be setting up this fuel watching task into Claude and ask, "Hey, how do I set this up and put the key in safely and securely?" And lo and behold, we get some instructions on how to do so. Here, it looks like it's getting us to just go ahead and put it in our environment variable. And I'm just going to be a sheep here and trust that Claude is pointing us to the right place. But if this was something important, I would be vetting this before I do it. You need your security thinking hat on a lot of the times when developing AI systems.

Sweet. Our key's in there. Let's go ahead and paste in that prompt and see what happens. We got some tool calls. Well, we got some writing to files and thinking, well, that's a good sign. Let's see what it does. And after a couple of minutes of trying and figuring out what's going on, we have an output. It's also gone ahead and, you know, identified some issues it ran into. You know, it's got a limited list of brands that it can be, you know, looking for fuel stations on and it suggested some fixes for it.

The development process from here is just going backwards and forwards and getting it to, you know, do stuff. Just prompting it in plain English. I also went ahead and asked it to set up a demo cron job to run this task you know two minutes from now and send it through our telegram channel and we got that through nicely as well and as you can see I did a test on our news channel thing as well less important you know you telegram discord whatever you use and run a test for it now there was a little bit of trouble here because I had actually initially set up my bot wrong and all I had to do was you know tell Claude hey something's going wrong here went through and diagnosed it ran some tests and once it found the issue I just said can you write a prompt for OpenClaw to let it know what was wrong and what to do in the future. And we saved that as some important info to reference next time we want to set up any sort of telegram communications. If your agent works through and tries to do something and you know it finds a solution, get it to save it. Get it like this is just it's like a different framework to be just thinking through. This is like it's sci-fi. Like it's so cool.

This fuel tracking really simple demo and it needs some more development, but we're leaving this as a nice end point, as an exercise for you to go off and do and just a starting point if you can't think of anything to do. This is a simple skill that could be expanded so much. You could be driving around the city for work purposes. You could be driving around the city or, you know, traveling between cities. If you can find a way, you know, if those routes are predetermined or pre-calculated and you know where you're likely going to be today, you now have a way to automatically find cheap fuel along those routes. Again, these are some really simple demos we're doing here just to get you started, but the range of what you can be doing with this is just mind-bogglingly large as there are so many tasks and niches that can be automated.

Maybe you set this up to read all incoming emails and check that it's been routed to the right department. Maybe you read a customer support ticket and fetch relevant data from a database to equip a human with what they, you know, need to be handed to them to get this ticket done quickly. Maybe once the ticket is figured out, it drafts a reply using the way that your company speaks so that a human can, you know, read over it. Yeah, that's good. Send it. Maybe you track quotes and follow them up after a certain amount of time. Maybe you keep this thing completely offline and you know plug in all your company documents and create a bot that can browse through the entire searchable history of your business. Which supplier has this valve the cheapest? Which has consistently gotten it to us quicker? Just go out and start playing with this.

However, you should do so safely. We can't speak for every use here in every situation, but keep your security thinking hat on when developing. You know, especially for a business environment. If somebody sends you an email, how are you going to protect against them? You know, putting a prompt in there to upload all your AWS keys to www.freesto.com. If you have your AI plugged into maybe secret business stuff, do you really want it to be automatically sending stuff off, you know, without human verification? It might accidentally leak stuff, you know, just have a human in the loop somewhere. Is that maybe a smart idea? Does an AI that can hallucinate really need the ability to write to your database or should it read only? Should it have, you know, very low permissions? If you're using external channels like Telegram or Discord, should you only trust whitelisted people? These are kind of the sorts of questions you need to be asking when developing this.

And I also want to end here by making you aware of something else here, and that is the fact that you can get OpenClaw to run inside a secure environment of itself. Now, it is mid-2026 and we can't predict the future, but right now, NVIDIA's Nemoclaw is pretty good. Essentially, it's a deny by default zero trust security layer that you can put over open claw. Want it to access a website, you have to poke a hole inside, you know, Nemoclaw and set up network policies to allow OpenClaw to, you know, peer out into the web, you know, the infinite web verse and be able to access it. Every new site that we listed there for it to check would have needed to be set up in Nemoclaw to give OpenClaw access to it. OpenClaw wants to access files outside of its own. Again, zero trust by default. You have to manually set all of this up.

While this is not strictly needed, it is maybe a good idea as it brings a bit of, you know, more enterprise level security to OpenClaw. It's not a complete replacement and solves all your security issues. It's just, you know, a little nice security cog in this machine. The reason why we didn't use it, look, I'm not going to sugarcoat it. It's not completely beginner friendly. It is something, you know, more professional and enterprise level and brings you, you know, enterprise level features, but the cost is that there is that extra moving part in all of this. You should at least check it out, though. It's easy to set up. You can just point Claude to the Nemoclaw page because there's even a prompt informing your coding agent how this all works and how to set it up as well as, you know, they've got tutorials there on setting up more OpenClaw agents if you're just interested in that stuff more. The prompt we used here was from one of their demos, so go and check that out.

I think my suggestion here would be to learn OpenClaw in, you know, a bit more of a safe environment, get a hold of the ropes and then maybe [clears throat] give something like Nemoclaw a go because you can just transfer over all those markdown files and skills and tools quite easily with Claude. It's pretty easy to migrate when it's just markdown files.

And that is where we are going to leave OpenClaw for now. The goal of this video was just to get you primed and thinking and you know just dangerous enough to go out and start playing around and developing this because we can't teach you everything that OpenClaw can do. This is like you know learning how to code. There is just so much to learn and so much to do and it's just you know really cool that you can get an AI agent to not only program and set itself up to do a task but then also go ahead and run it you know when it's called.

And if you are serious about this, we are going to do a nice little upgrade to this system in a very near future video with LiteLLM, that model selector bit of software that we're going to put in our stack. And if you made anything cool with this and you want to share it or you just need a hand with anything that we covered in this video, feel free to head on over to our community forums and post about it. We're all makers over there and we're happy to help. Until next time though, happy making.