Local AI Agents In 26 Minutes

Added: 2026-05-04

[00:00:00]I learned all about local AI agents for you. So, here is the cliff notes version to save you the months that I have spent playing around and building with local AI agents. Like Open Claw, Nemo Claw.

[00:00:10]Look, here is my little 24/7 Claw Factory during research and building me software. So cute, right? And also no code safer options like Anthropic's Claude Co-work. Honestly, I waited many moons to actually make this video cuz I wanted to be 100% sure that this is not just like a hype thing, right? I am now 100% convinced that local AI agents is a new category of AI products. That honestly in the past few months has high-key transformed my daily workflows.

[00:00:35]Now, without further ado, let's go. A portion of this video is sponsored by Grammarly. All right, here is the outline of today's video. First, we're going to define what local AI agents are, including the anatomy of a local AI agent and how to custom design your own.

[00:00:48]This is where the majority of people fail to get the most of their local AI agents cuz they don't actually understand how to design them properly.

[00:00:54]I'll also cover some principles and tips including how to deal with safety and good engineering principles. Then we shall make this very concrete. I will be showing you demos of code and no code versions of local AI agents, specifically Open Claw and also Claude Co-work. And finally, I will end with some thoughts on how to take advantage of this new category of local AI agents.

[00:01:13]Because this is just the beginning and represents massive opportunities. Let's now actually define what a local AI agent is. First and foremost, it is an AI agent, which is an AI that can take actions and complete task on its own. AI agents are nothing new, they've been around for over 2 years now. But what makes it really exciting and more powerful these days is that it's also local. Local meaning that the agent is able to live and run on your personal machine. So, a local AI agent is an AI that takes action and completes task on its own while running directly on your machine. So, it can do things like send you a very personalized morning brief of your calendars, your emails, priority stock portfolio, notes, and news that you're interested in. For me, my local AI agents are also constantly researching ways for me to improve my business and then autonomously build out software to implement these things too.

[00:01:57]For example, my local AI agents have built me a personal finance tracker, accounting software, a dashboard that tracks our boot camp students and their performance and their progress, and some trading bots. Not financial advice, it's just a personal hobby I have. You can also remotely interact with your local AI agents like communicate with them, monitor what they're doing through your phone as well. So, you don't have to be sitting there at your computer. There are so many different use cases and I'm just going to put on screen some of the most useful ones in my opinion. So, take a screenshot when you're ready to build your own later. Now, I want to show you the components, the anatomy of a local AI agent so you can custom build your own. So, picture this little guy is your local AI agent. Its name is Inky and we are now going to build and customize it.

[00:02:35]The first thing we need to do is decide where it lives. We know that it's a local agent, meaning that it lives on your machine. But there are a lot of different types of machine it can live on like your personal laptop, a old laptop, a Mac mini or Mac Studio, a PC, or even a VPS, which stands for virtual private server. It's like renting somebody else's computer on the cloud.

[00:02:54]In my opinion, there are three key factors to determine where you should be running your local AI agent. The first one is if you want your local AI agents to be continuously running 24/7. So, you would need to have something that's running 24/7. So, something like your laptop, which you carry around with you, wouldn't work. Then you need to consider your machine specs, things that what kind of chip it has, CPU, GPU, but most importantly, RAM memory. Because if you want your local AI agents to be powered with really big open source models, we'll get to that in a little bit about different models. But basically, you need to have a machine that's able to handle running these models. So, something like your everyday computer probably wouldn't cut it. third factor is privacy security. How paranoid are you about your local AI agent having access to your things? Unfortunately, there have been stories of people who installed their local AI agents on their like personal laptop, not put in the right guardrails, and it has like access to all of your stuff and it like deletes your emails potentially, could like do not great things because it has access to everything.

[00:03:51]So, if you're very particularly paranoid like I am actually, then you probably want to have a machine that is just dedicated for your local AI agent and it doesn't have access to all of the stuff you don't want it to have access to. So, me personally, when I first started playing with Open Claw specifically, I went with a old laptop like this one here that I completely wiped so there's nothing on there anymore and I have it on 24/7 completely dedicated to my local AI agent. Because this old MacBook Pro only has 16 gigs of RAM, I mostly just use Claude's Sonnet and Claude Opus models instead of big open source models cuz that wouldn't be enough RAM. Again, I'll explain a little bit more about models later in the video. But yes, after playing around and building stuff more seriously, I also started using a Mac mini, which is a dedicated machine that also has better hardware. So, then I can start running bigger open source models too. I also have one on a M2 CS and I actually just ordered a Mac Studio as well because I'm like really into local AI agents. So, yes, that's just my setup, my hardware machine journey.

[00:04:44]There are a lot of other options. I'm actually going to put on screen some hardware and machine specs and what kind of models and what kind of AI agents you can be running. And I'm actually going to put a prompt in description as well that you can paste into your favorite AI chatbot like Claude or ChatGPT and it'll actually help you figure out what kind of hosting option is the most suitable for you. I got you. I know like hosting and hardware and stuff like that is a really big blocker for a lot of people to play around with local AI agents.

[00:05:08]Okay, great. Now that we've talked about all the different places that our local AI agent Inky lives, let us now actually customize other parts of it. Let's start off by giving your local AI agent Inky a mouth and ears, aka a communication channel. So, you know, you can actually talk to Inky and Inky can talk back to you. This is also what allows you to talk to it remotely with your phone as well. So, there are actually a lot of third-party communication options. Many of which you probably already use and know of like Telegram, Discord, WhatsApp, iMessage, Slack, and Dispatch.

[00:05:36]The easiest one to get started with is just single channel messaging and Telegram is usually the option people go for. Or Dispatch if you're using Claude Co-work. Then as time goes on and you get a little bit more advanced, you got like multiple things going on with your local AI agent, you probably want to graduate using something like Discord that has multiple channels there that you can stay more organized. I will actually show you what this looks like in the demo later. But just for now, know that there are a lot of different options available. Okay, great. Our local AI agent Inky now has a mouth and ears, communication channel. Next up, we need to give it a brain and memory.

[00:06:07]Let's first start off with the brain, which is the large language model, the AI model that would power it. There are a lot of different options you can choose from like Claude models, Opus and Sonnet, Open AI models, and of course open source models too like Qwen and Kimmy, MiniMax, DeepSeek. Different kind of model have different kinds of tradeoffs like capabilities, size, speed, cost, and privacy. I actually made an entire video, which you can check out over here, that goes through all the different types of models out there. But TLDR, the most popular options for local AI agents is Claude Opus or Claude Sonnet. And on the open source side, Qwen and Kimmy are the most popular. I'm going to put on screen now some different types of models and what they're commonly used for, what they're best for to help you make a choice on which one you want to use. So, take a screenshot. Okay, great. So, the model is your brain. But you also have to give it memory so it can know things about you and can remember what it is that it's doing. This is probably a lot simpler than you would expect. Memory is literally just a bunch of text files and you just dump in there document everything that you want Inky want it to know. Stuff about it, who it is, what personality type, what it's supposed to be doing, what kind of workflows it has, what kind of data that it has, and information about you, the user. What's your workflows? What's your personality type? What do you care about? What's your demographic? What's your job? This is what gives it ability to be very personalized to you and what you want it to do. And as it's completing a task like I don't know, doing some research about stocks, it would be writing down what it's actually doing. So, in the future when you talk to it about these stocks, it would know what you're talking about. The good thing is that most local AI agent software and frameworks like Open Claw and Co-work do already have some type of memory system that's pre-built in it. But people who are like power users do like to do things like boost the memory system to make it more robust and organized by using things like Obsidian for example.

[00:07:42]I will also be showing you this in the demo later. Great. Let's now give your local AI agent Inky some more tentacles.

[00:07:47]Tentacles represent skills and tools.

[00:07:50]The good thing is that Inky it already comes with some pre-built tentacles so it can already do some basic things like search your files, execute code, things like that. But depending on what it is that you want it to do, you want to give it more abilities, more tools that it can use, and more skill sets that it can have. This can include things like web search, being able to access your email, take screenshots, text-to-speech, image generation, etc., etc. Wow. Okay, next up, let us give our local AI agent Inky a heartbeat. What it is is that allows you to schedule task so that Inky would run without you asking it to do so. Like every 30 minutes it's supposed to scan for new email, every month it should be scheduling a doctor's visit. Or it can be time-based called a cron job. So, every morning at 7:00 a.m. it sends me a morning briefing. It can also be event-based like every time a file gets added to my accounting folder, it would run my accounting workflow and put it into the books. These are such game-changers. I'm going to put on screen now some of my favorite ones.

[00:08:44]Yay. Inky now has all of its body parts except one last thing, eyes. Having eyes would allow your local AI agent to be able to see what is actually happening on your computer and interact with it like a human would. So, it can do things like check what's in your folders and in some cases actually be able to see what's happening on your screen and be moving your mouse and actually doing stuff on your screen for you. Here are some other examples that Inky would be able to do if it has eyes. All right, we have now finished building Inky. Inky has all of its body parts now. I'm going to put on screen now a little diagram that shows the anatomy of a local AI agent. Take a screenshot, which will become immediately useful because time for a little quiz. Please answer the questions displayed on screen now to make sure that you are retaining all the information. Now that you have your custom little designed Inky, your custom local AI agent, there is actually no reason for you to only have just one of these. You can actually have multiples of them. You can get your little local AI agents to be doing multiple tasks at the same time. Or you can put them into teams where each of them has a specific function and added together they're able to produce something that is greater than the sum of its parts. For example, I have a research team that is looking into stocks that I might be interested in buying. And I also have a software team that is researching and making product decisions and building software for my company. They can be doing multiple things at the same time. One of them is doing research, one of them is managing your calendar, one of them is writing a marketing campaign for your job. So, there are specific principles for designing these multi-agentic systems, which I'm not going to go into too much more detail in this video cuz it's There's like a lot going on there as well. But, I actually did make a video where I covered multi-agent system designs, which you can check out over here. As a YouTuber and entrepreneur, my workday is all over the place. One hour, I'm deep in the codebase. The next, I'm writing a video script. And then, I am writing a LinkedIn post. The writing never stops. And the context switching, getting started on things, the cold start problem, is brutal. So, that is why I use a Superhuman Go to stay on top of everything. Superhuman Go, which is from the makers of Grammarly, is an AI partner that works inside the apps and websites that you already use.

[00:10:47]Proactively making suggestions to ease your work without you even asking. Let me show you how I use it. First, turn on Superhuman Go within your Grammarly browser extension and select the Go icon on the side of your screen. And it just shows up. If I'm looking at a LinkedIn post draft or a creative brief, I can ask Go to summarize it or help me understand exactly what's being asked without opening a separate tool. I really love that it just shows up wherever I'm working and I don't have to open up a separate tool. My favorite feature is reader reactions. As a creator, I am always wondering, will what I say actually land? Well, reader reactions predicts how your audience will respond before they even see it. It flags where they might get confused, what questions they'll ask, and suggests revisions to make my writing more effective. I use it on everything from scripts to client emails. You can also connect external agents like Google Calendar or Slack directly inside Go.

[00:11:31]So, everything lives in one place without having like five different tabs open. So, if you're professional juggling a million types of writing everyday, this tool is for you. Try Superhuman Go to level up your productivity at work. The link is in the description. Thank you so much Grammarly for sponsoring this portion of the video. Now, back to the video. Okay, so before I go into demos now, I do also want to cover two final principles that I think is very important as you are building and working with your local AI agents. The first one is the safety.

[00:11:57]Safety is the primary concern for using local AI agents because you're basically giving this very intelligent agent access to your computer and hoping that it's not going to just go bananas. So, that's why you need to take precautions to make sure that your local AI agent doesn't just like ruin your life. There unfortunately have been stories of these of local AI agents deleting people's emails or having viruses in them because of like skills being shared, things like that. So, yes, always keep safety in mind. And the way that I think about it is first of all, I try to isolate uh the local AI agent as much as possible, which is why I don't run them on my primary machine that actually has all my data. I only run them on dedicated machines where there's no sensitive information that I can have access to.

[00:12:36]I'm also very careful about what I give it access to. Like, it does have access to some of my emails where I want it to be screening emails, but I don't actually give it access to my personal emails, the ones that actually has like very sensitive information on it. I use other emails for that. I create new emails. And I also don't trust other people's workflows. Like, people would write workflows that document doing a specific process, like writing a marketing campaign, right? And these are like really useful, but there could also be like malicious things that's put into it where if my local AI agent gets access to it, it might cause it to go bananas. So, I generally just don't use other people's skills unless it's from very trusted developers. And if I do want to use a skill I think it's like actually a really good skill that somebody else made, I would actually give the skill to Claude, I'd tell it to scan the skill, and then rewrite it itself, and then give it to my local AI agent. General rule of thumb is just to be as paranoid as possible. And finally, local AI agents, like if you're using Open Claude for example, they do have like dedicated security things that it tells you to go and check. So, what I actually do is use my AI agent's heartbeat ability and run a security audit like every hour or at minimally every day. Now, if you do choose to use something like Claude Co-work for example, the good news is that a lot of these security functions are already pre-baked into it, so you don't need to worry as much. Now, the second principle that I would like you to keep in mind is good engineering principles.

[00:13:52]Specifically, to always be giving clear instructions as much as possible for what it is that you want your local AI agent to be doing or building. And only adding one feature or workflow at a time. So, it's easier to track and monitor. Don't be like, "Hey, Inky, build me like five things at the same time." No, cuz then if something does go wrong, you don't even know what's what's up, you know, you don't even know what's going on. Very chaotic. All right, it is time for demos. I'm so excited. Starting off with Open Claude. I want to show you guys how I implement all of these components that we just covered and the custom stuff that I'm doing with it. All right, cool. Time for the Open Claude demo. So, here we have the agent office where you can visually see all of the agents and what it is that they're doing and how they're interacting with each other as well. Uh by the way, the visual here, this is what you're seeing here is a custom tool called the mission control. It allows me to monitor what the agents are doing and then also like other tabs as well, which I will show you guys in a little bit. But first, let's talk about the team that's here.

[00:14:47]So, here is the mission of the team, what everybody is working towards. And there's me. So, I'm the founder and CEO and I'm also the human. There's Inky, who's chief of staff. We can see that Inky is using the Claude Sonnet 4.6 model as the central brain. Then, we have a Dinky content pipeline. So, that's Blinky, Pinky, and Dinky. It's an autonomous content pipeline in which, yes, I will show you in a little bit. I think it's pretty cool. These are all using the Sonnet models, too. Then, we have Linky, which is the builder, decoder. So, every time we need to build software, do stuff with code, that would be routed to Linky. So, Linky has two models that it uses. The Claude Opus 4.6 model, which is for planning, architecture, and more like intense coding stuff. And also has QuenCoder 2.5. This is an open-source model that actually lives on the computer, so it's completely free. And Linky would route any more mechanical coding task to QuenCoder because that's a way to optimize costs, right? Cuz Claude Opus is really expensive to be running all the time. Then, we have Winky, which is a system monitor. It runs twice a day and it just basically does a health check to make sure all systems are up and running, there's no security issues.

[00:15:48]And it uses the Mini Stroll 3B model, it's also a very small local model that is free in order to do this. By the way, all of this, all of the agents and then the models as well, this is all on a secondary laptop that I have. So, that is the home for this Open Claude setup.

[00:16:03]I have many Open Claude setups, but for this one, it's actually all just running on a MacBook Pro laptop with 16 gigs of RAM. I'll actually put on screen the stats for this MacBook if you are curious. So, let me actually show you the full content pipeline. So, this is actually done on Discord, which is the communication channel that I use to interact with all of the agents. There are different channels that represent different things it's doing, current projects, and different alerts that I get from different agents as well. With the content pipeline works is that I get a morning brief that gives me the top stories of things I'm interested in, which is basically like AI stuff. And then, it goes and updates my topic watchlist. So, this is how it is that I actually track different topics to decide when I should make a video about it. And it also gives me some video ideas as well. It would also take these video ideas and put them into the content ideas channel here. I look at these content ideas, I'm like, "Hmm." If I like any of them, I can go on my mission control and look at the content tab. I look at the content ideas that is also over here. So, say that I like something, AI agent memory explained.

[00:16:58]Like, "Oh, okay, I'm going to actually do this one." So, I can click it and it goes into my content ideas. And then, I can click approve or pass. So, in this case, I have approved local AI in N minutes and also AI infrastructure explained. And we can actually see here it routes to the YouTube long-form channel and it gives me a working title, the topics, and my ideas, and some bullet points about it as well. And then, I'm able to get the video outline for it. AI agent memory explained in 28 minutes. It would give me um some bullet points about how it is I can structure the video, etc., etc. Of course, I still need to actually flush this out myself and like cross-reference, check a bunch of things. Takes me a very long time to actually make a video, but this is a really good starting point. So, to do this, uh Blinky, Pinky, and Dinky of course need to use a bunch of tools. So, we can actually ask, what tools are being used for the daily digest to content pipeline. And it tells us we have web search, external, etc., etc., TLDR. These are the tools that we're using. Another way to get different types of tools and skills is to go on Claude Hub over here, which is where people share different skills that you can download and use as well. Like, self-improving agent, ontology skill better, etc., etc., GitHub, weather, blah, blah, blah. Me personally, I don't actually download things from Claude Hub unless it's like literally by the founder, like Pete Seinberger for example, because there was this whole like security risk and scam situation through Claude Hub previously. So, I'm just being paranoid. But, what you can do is I say you like this, you can click on it and then you can actually just give this to Inky and just tell it to build the skill itself without having to download it directly. Cool. So, here is a board that shows all the tasks that we're doing, what I'm doing, what everyone's doing, what Inky's doing, what Tina's doing, as well as an activity log. Here are some of the jobs that are being scheduled and the projects that I'm currently working on as well. Oh, another thing that I asked my Inky to do is that every night it would build something by itself that is delightful for one of the current projects that I'm working on. So, when I wake up, I'm able to like see something new that I built. Like, for example, last night it built what AI skill should you learn next? Take an MVP for a micro SaaS. And it's just like this little quiz that it built that's able to help you determine like a personalized AI learning roadmap. Am I going to use it?

[00:19:07]Am I not? I don't know. Uh but, it's really fun waking up every morning to see something that gets autonomously done. And sometimes I actually do like the idea and I start building on top of it, too. Another thing is memory. So, I do get Inky to aggressively document everything it does. So, here you can see all the docs that it writes for everything that it builds. And you also have the daily logs, everything it's doing every single day, and its long-term memory as well, which is memory.md. And since Inky has eyes, one of its task every day is also to see all the changes that have been made and then aggressively document everything to make sure that nothing is left undocumented.

[00:19:40]I also have this in Obsidian here. So, you can also read it through like Obsidian if you want, which is low-key pretty cool. I'm probably also going to work on this to develop it into a more robust memory system that would also function as a second brain for myself.

[00:19:51]So, yeah, this is my current well, one of my current open claw setups. There's a lot more I can talk about, so let me know if you want me to make a more dedicated open claw video. I am happy to do so. So, cool. That was open claw. I also want to show you a safer no code option, which is Claude Co-work. So, Claude Co-work is also a local AI agent and it's Anthropic's take on local AI agents. It's a lot safer and it's a lot easier to use, so I really recommend it for people who are not comfortable at all with code or just starting out. But, of course, the downside is that you do get locked into Anthropic's system and there's less ways for you to customize things. But, still, I think it's a really good option if you're just starting out. Welcome to the Claude Co-work demo. So, here we have Claude Desktop and this is the central hub for Co-work. You can see the different models that are available over here.

[00:20:35]Obviously, they are all Anthropic models and you can tell it stuff. Hi. The interface, as you can see, is very similar to how you would usually chat with a chatbot and you can chat with it on mobile through Dispatch as well, which I will show you guys in a little bit. But, first, let me explain how all this works. So, Claude Co-work lives on your computer and it basically corresponds to a folder on your computer. In this case, I'm using my personal laptop because I do trust Claude Co-work more than I do Open Claw and like not revealing all my passwords and destroying my life. So, yes, it's over here and it also has these memory files, which is claw.md and memory.md, which I like to look at through Obsidian. So, here we can see that information about me. This is Tina's Co-work workspace, blah blah blah, memory system and things about me. So, going back to the file system, I've made separate folders for different projects like content studio, portfolio, lonely octopus, and personal. So, why don't we look at portfolio right now? This is my investment portfolio, by the way, and it has a portfolio dashboard. So, say I want to ask questions about my portfolio. I can choose the project portfolio folder and I can ask it some questions like, "What is my top We'll use the Sonnet model and we'll say, "Let's go." Cool. And it tells me that, apparently, my best performing winner is on the Hong Kong Stock Exchange. Pop quiz, does anybody know which stock this is? Not financial advice, okay? Please, we please not financial advice. In fact, I'll also show you what's my biggest loser. It's also lose, too. There you go. My China AI names are all underwater. Got to stay honest. Now, that's cool and stuff. You can build out projects, but what's really cool is that you also have access to Claude Code, which is also on this hub. For example, one of the things it built me is this investments dashboard.

[00:22:06]So, it actually shows me some of my investment portfolio information is about it and it updates by itself as well. Has some positions, research about segments I care about, and a watchlist that I have here as well. The way you do this is, of course, through a combination of Claude Code. You also have access to other types of tools, including skills and connectors. Not going to go into too much detail about what these are, but basically, you can have lots of different apps that you can work with. So, say you want to pull from your Google Drive, Google Calendar, you can do that and also have automated workflows. Plugins are basically combinations of skills and connectors for a specific purpose. Like under finance plugin, you have skills like different audits, close management, etc., etc. And under connectors, you have different apps like Google Calendar, Gmail, Microsoft 365, etc. There's so much that I can do with this as well. Don't have time to talk about it right now, though. Let's see. I also want to show you the scheduled tasks, which is the heart of your AI agent. So, in my case, some of my scheduled tasks would include every day at 8:00 p.m.

[00:22:59]doing a daily investment deep dive, portfolio daily briefing, and daily macro briefing as well. This is all reflected in the investments dashboard that I built. Okay, so finally, I want to show you guys Claude Co-work's eyes, aka computer use, which is arguably one of the most visually cool features from Claude Co-work. And I'll be doing this through mobile. So, say, for example, I'm out and about and I need a file on my local computer that I'm out and about. So, I'm only on mobile. What I can do is actually go on mobile and ask, "Can you send over the Discord PNG file?" I know I want this file. Here.

[00:23:33]Yes.

[00:23:35]Click send. It will do its thing. Can allow it permissions and is able to send it to you. And you're able to on mobile have access to it now and do whatever you need. So, it didn't really do computer use here. Let me try again. Can you take a screenshot of it? And yeah, it's literally using your computer and doing stuff with it. So crazy. It's searching.

[00:24:00]Opens it and it sends you a screenshot, which you can see over here. So, yeah.

[00:24:06]>> [laughter] >> Pretty cool, right? There's, of course, so much more that you can do with a Claude Co-work system as well. I personally actually use Claude Co-work and Open Claw at the same time and like some other agents, too. Like many different local AI agents for different use cases. All right, final section. I want to talk about some predictions about the development of these local AI agents and the things that you should be learning to take advantage of this new category of AI product because it represents massive opportunities. It's not just me that's saying this, by the way. Check out this clip. This is Jensen Huang from Nvidia. The implications incredible. First of all, the adoption says something, you know, all in itself.

[00:24:42]However, the most important thing is this. Every single technology company for the CEOs, the question is, "What's your Open Claw strategy?" He's basically saying that companies all need to have an Open Claw strategy now. A personal local AI agent strategy. Every single SaaS company will be coming in a gas company and a gen tech as a service company. As he explains, there are so many opportunities out there for your personal workflows and for your business workflows. My recommendation, if you're someone who wants to take advantage of this rise of local AI agents, is to actually learn to build your own AI agents, too. Like, yes, play around with Open Claw and set up these different local AI agents, but if you understand how to also build your own agents, this is so much more powerful, especially in combination if you learn how to do AI coding. Like, oh my gosh, if you know how to use these local AI agents, you know how to build your own agents and set up systems for it and also know how to do AI coding so you can augment these local AI agents, too, that combination is so overpowered. Like, the world is your oyster. All right, so that is the end of this video. As promised, here is a final little assessment. Please answer the questions displayed on screen now to make sure that you're retaining all the information that we just talked about and you're ready to start doing stuff with your own local AI agents. Thank you so much for watching until the end of this video and good luck and have fun. I will see you guys in the next video or livestream.