How To Setup Claude Flow To Run Agent Swarms! I CAN'T BELIEVE IT!

What? Oh my god. Oh my god. It worked. But the real thing, it's time to move on from using Chad GPT to help you program. So if you follow along with this tutorial, you can build something like this. In this tutorial, we're going to be looking at using clawed flow on top of clawed code to have an agentic swarm go off and build an application for us. So, this will be for users on Windows using PowerShell in a WSL setup. So, with that said, let's jump into it. To kick things off, I'm over at the GitHub repository for Claude Flow. And Claude Flow is going to be what we're using to basically have an agent swarm over top of Claude Code. The way that this works is you have a couple of different modes to run in. And there's many more features with this that I'm not going to be going over in detail, but this is essentially going to be set up for how I use this. Claude Flow allows you to run a swarm of agents or you can run what's called a hive mind. And there is a slightly different architecture in terms of how these agents are set up. But with the hive mind, one of the things that's really cool is that you sort of get this shared memory between your runs of these different swarms. It's cool in that, and you may have noticed this when working with agents, you're trying to almost like teach it like, you know, don't run this command or always run these things and you're updating either your GitHub copilot file or your claude file, whatever it happens to be with more instructions. But this shared memory idea is a nice way that you have this like learning mode for your agents. What we'll be looking at is getting this set up to run on Windows. And I'm doing this in a particular way because I've noticed that up to this point, Cloudflow has been a little finicky on Windows. They are supposed to have first class support for it, but I've just found that it's not really working how I would expect. What we'll do is we're going to be using Windows. I'll be using PowerShell. And then we'll be using WSL, so the Windows subsystem for Linux from PowerShell. And then we're going to install these things, have them run, and we'll build something cool that hopefully works. From here, what I'm going to do is jump over to the install WSL page. This is from Microsoft. Just to explain again, we want to be able to have a Windows subsystem for Linux running from PowerShell. So, if I scroll down a little bit here, you can see WSL install. I already have this installed, but all that you would do is simply run WSL-install from PowerShell, and you should be able to get the walkthrough for getting WSL. And if you'll see that if I just run WSL now, notice how my terminal has changed, right? I am now in the Windows subsystem for Linux. And you can tell like purely just because of this here. This is very much a Unix type of thing versus here. This is very much a Windows type of thing. And if I type exit, it will take me out of that. But we will want to be in WSL. For me, what I'm going to do next is navigate to where I would like to be creating my projects. I have a development folder. So, I would go over to that. So, this is where I have all of my code that I uh building for projects or things that I put on YouTube, that kind of stuff. But, we're going to install what we need sort of at a global level here. So, it's not going to be like every single project you have to go doing this, but you will have access to this stuff. So, give me one sec. We have WSL installed, but I'm going to scroll down a little bit lower on the GitHub page here. We will need to get Claude Code installed. So this is going to be sort of the very first thing that you need. So copy that. Again, I already have it installed. So you want to make sure that you're installing it. The - G flag is for global. And then make sure that you're doing this because if you don't, you won't get the right package. And I have made this mistake because I go to install it and nothing actually happens. But just to prove that it's installed, if you type claude now, and I already have it, like I mentioned, you should be able to see what looks like this type of interface. And if you navigate into it, you will want to proceed. And then from here, this is clawed, but if this is your first time getting into it, you'll want to set up claude to make sure that it can have the either the free mode or you can set it up so that it's using your account. And that way if you're, you know, paying for a Claude subscription, you can have the right limits and stuff like that. So I'm just going to exit that. So now we're here. So Claude is installed, but Claude Flow at this point is not. So let's head back over and see if they have the actual instructions. And it might be probably somewhere along here, but I'm just going to show you that it's pretty simple, right? So it's going to be the same type of thing. I'm just going to copy and paste that. and you would do COD flow. And then I'm using the alpha version right now. I don't know what the release cadence and versioning is going to be for this, but the alpha right now has a lot of the newer features, especially some of the fixes that are supposed to be coming to Windows. At the time of recording this, and probably over the past couple of weeks, the Windows support for this has not been stellar. And I totally believe that this will be improved, which is why I'm sticking to this. But I noticed that some of the shared memory features are just kind of broken. And that's one of the biggest things that attracts me to this is because I spend a lot of time trying to guide AI agents and you know I'll have a conversation with AI it's doing something and I'm like hey man don't do that clearly that's the wrong thing and it does the typical oh you know of course you're right but then of course later in the conversation it repeats itself and I'm going like please I wish you could learn and um it should be able to but um it's a little bit of a pain in the butt right now and then between different swarm ARM runs. It's supposed to keep that shared memory, but it's not persisting currently on Windows. So, fixes are coming. But you would run this npm install -ashg for global claude flow and then alpha. And now I have claude flow installed. So, if I run cloudflow, right, you can see that I get all of these commands here. And I'm going to show you now how I go to set this up for a new project. Okay. So, it's all installed. But what I'm going to do is make dur and then um let's call it and then I will go into that. Okay. And now we need to get claude flow initialized for here and with the hive mind setup. So that looks like this cloudflow init. And then there's this option for what's called spark. And at the moment, Spark, the acronym, is uh evading me for what it stands for, but it's essentially going to set up a bunch of uh documentation files and instructions for you can see like Spark methodology right here. So, uh it basically guides the agents to like a specificationdriven design, which is really cool. You may have noticed with agents that they do a lot better when you give them a lot more specification upfront and more structure and guidance. and this kind of forces them down that path. Something that I'll mention from my own experience, I'm sure this kind of stuff gets better over time, is that with agents, if you've used one agent before, sometimes you might feel like it's doing cool things, but kind of feels like it's sometimes not that smart, like it's making really obvious mistakes. And when you bring a swarm together, it's kind of that problem gets multiplied. So now you have more agents kind of making silly mistakes. So more specification and guidance up front really helps a lot. So I'm going to put this here as well. And you can see that it's copying all of these markdown files in. Right. I'm just scrolling up a little bit so you can see what's going on. Right. So it's initializing it. Get all these things set up. It will talk more about I think in this version it does say that it's having issues finding Claude. Right. So this is a current bug. It says it can't find cloud code CLI. It's I already showed you it's installed and this will work. So if we continue on, it will still work. It seems like this is a bit of a glitch in the initialization. And also if it's telling you this because your install is not quite right, then you would want to go run these MCP servers. But if I were to copy and paste these lines right now, I might as well do it. It will say that they're already running. So let's go do that. Oh, in this case it wasn't. So I think when I've done this before, it's already been updates to existing directories. So cool. Good thing we did that. Now we have our MCP servers configured. Maybe a previous version of this already ran them. I don't know. So maybe I've been used to that. But you can configure more MCP servers. If you go back to GitHub, you can follow through the documentation. It will guide you through how you can do that. I have not had to set up anything custom yet for my own development. So, this has all been out of the box and worked well for me, aside from the shared memory part, but fixes are coming. Now, we have that. The next thing we want to do is get the hive mind part set up. So, cloudflow hive-mind in it. We'll go run this. And you can see very quick to get started here, but created the hive mind directory. This is the part that is currently not working unfortunately. But what will happen is that the hive mind is supposed to use the SQLite database as the shared memory. Currently on the Windows versions, it makes it look like it's working, but it falls back to inmemory. And that way when things end, you kind of lose all of the memory, which kind of sucks. But next part, there is two ways to get started here, and they they shut they tell you both. One of them is just to spawn a new hive, and the other one is to go through the wizard. I haven't checked recently if the wizard is fixed. I personally like this is in a terminal. I'm a a user interface person. So going through a wizard for me is like heck yeah that's my preferred way. But when I ran this before the wizard just was not working. It would go create the hivemind and then it would never run again. That could just be a Windows problem. Maybe that's fixed by now. But we're going to walk through this version here. This is what the format's going to look like. I like putting clawed at the end and then auto spawn. I noticed and I'm just giving you these details because at this point in time, if I don't include these, it doesn't actually start properly and I don't have another way to start the hive mind once it's created. If I use these parameters that I showed you, it will kick off automatically. So, take that with a grain of salt. By the time you're watching this in the future, this may be corrected behavior. But, there's something cool that I like to do now, and I'm going to demonstrate it by having my hands visible so you can tell. But what I've done is I've purchased a foot pedal and of course I'm using a microphone. But now that I'm working with agents a lot more, I've been finding that trying to type the things that I want to agents is a bit of a pain in the butt. And that's because a lot of the time it's doing things and then I have to like stop it and I'm like, "No, no, no. Like you must go do this." And what I'm trying to tell it a lot of the time it would kind of be like trying to tap a friend on the shoulder or a colleague or peer that you're working with like to tap them on the shoulder when you see they're doing something wrong and then trying to type a message to them. It feels really really awkwardly slow. So I invested I think it was like 50 bucks or something in a foot pedal. It's got three switches on it. And just to kind of show you, if I press, you can't see my feet, but if I start typing like this, and I'm using Whisper Flow, which is another AI tool, and if I stop on the pedal, you can see, right, my hands are still here. Um, and it records what I'm saying. So, I found this has been a tremendous performance improvement for myself. What I'm going to do is ask this thing to go build something for us, and we'll see if it works. I'm going to ask it to build a blazer Pokédex. >> Venusaur, the final form of the Bulbasaur evolution. >> Just a quick reminder that if you're finding this interesting, remember to like the video and subscribe to the channel for more. We'll see what it can do. I would like you to build a Blazerbased application that resembles a Pokedex. I would like to make sure that we can track which Pokémon we have seen and which Pokémon we'd captured. And this must be int 9 as well. I don't know. Maybe we can start with that and see what it does. This isn't going to be a full build it from scratch tutorial or anything like that. I just want to show you what it looks like to have the Hivemind run after we've set things up. So, I'll press enter on that. And you can see that it's now kicking things off. So, there we go. Begin hive mine execution. This will basically print out every time that you're doing one of these, but the beginning of it is basically getting set up to figure out which agents to run. So you can see researcher poke pokey researcher. Uh it's a coder, an analyst, a tester, and then you can see that it's using memory, right? Memory usage and then trying to store this kind of stuff. I think that some of it does get stored in the SQLite database, but at other times when it comes to reading it, it's not working. So, what we'll see now is that it has a to-do list. And this is a really critical thing that you may have noticed for yourself when working with agents, but having a to-do list is really valuable because it keeps the agents on track. I find without it, they can kind of go off on their own and it just gets really crappy really fast. But if it has a to-do list that it keeps up to date, it should be a little better. That scrolled really high up. So, it's basically running through this now. I've just scrolled up a little bit higher so you can see the to-do list and I'm going to read through it as it's trying to do some work here, but it says research and planning phase. It's going to look for Pokemon APIs. Cool. Um, if it can find Pokemon APIs, I think that would be awesome. I'm not really expecting that to work to be totally transparent. Blazer patterns make sense. And this is constantly refreshing. Sorry. So, it's kind of scrolling all over. But um project setup of course the architecture design. I'm not really expecting a lot from the architecture here because I didn't really give it much guidance at all. Core data layer UI components state management. I think the thing I'm most concerned about is that it's looking for this like a Pokemon API. And if it doesn't find it, who knows what it's going to do. So just to scroll back down, updating to-dos. So it's actually making progress through this. It's gotnet found. I'm going to let this run and we'll see what it does. Okay, I never thought that we would get here, but we did. And hopefully my video editor made this a lot more enjoyable for you than it was for me. So, it has finished apparently. So, Pokedex Blazer application complete. I've successfully built a fully functional Pokédex application using Blazer Web Assembly and .NET 9. Here's what was accomplished. So, we have a Pokemon browser search and filter. We got a tracking system. We got detail modal. So, we can see comprehensive stats. Okay, very cool. We got loading start with 151 Pokémon load more on demand responsive and offline storage. Um, it talks about the architecture for this, which I was watching it talk about as it was all kind of blasting through and updating in the terminal. We have the ability to run it. It's going to be a localhost 5000. Okay, moment of truth. Okay, I'm going over to the solution folder that it created. And first thing that stands out to me is I I don't even see a solution. Um, if I go back up, is it here? No. So, there's there's no there's no solution. So, let's uh let's do it this way. Probably shouldn't have done that. One sec. Okay. new terminal. We're going to go into WSL. Then we have to go into the project file or the project folder. I don't understand how this is going to work if there's no solution. But oh, this is not going according to plan at all. All right. So, I'm going to go into Bash here. We're going to go into the folder that it made where there's no solution. And then we're going to Okay, let's see what it does. Apparently, it's running. Oh. What? Oh my god. Oh my god, it worked. Okay, so I wasn't actually expecting that to work at all. Now, when I say it worked, I don't actually know how much it's going to work. So, it says it's loading the Pokemon data. To be totally honest, I don't actually know what it's loading. I haven't looked at any of the code yet, but kind of cool. Peek. How do you use Pikachu? Can I search that? I don't know. I don't even think this is working yet because it's not actually loaded any data. So, this might be completely broken. It probably has something behind the scenes. It's like, go implement this with the real thing. I can't believe we just made a Pokedex. I can't believe it. We did it. We did it. Okay, so there you have it. We were able to get our Pokédex put together and that's all thanks to being able to use Clawed Flow using an agentic swarm specifically in a hive mind configuration. We were able to give it a simple command to go run. So this again was running on Windows from PowerShell in a WSL setup. Thank you so much for watching. I hope that you can go off and build some really interesting things with some help from your agent swarm. See you next time. I actually can't believe we got a Pokédex. This is the craziest thing I've ever seen.