Transcript: Ten new AI-development repositories: what is genuinely useful, and what is hype

Source video ID: LrHY1U03IRU

Transcript

• 0:00 — So this month we have seen a ton of brand-new, really cool open-source repos hit the scene. And in today’s video, I’m going to show you 10 of my favorite. And these range from skills you can add to your next Claude Code project all the way to completely customizable AI agent harnesses. So, whether you’re someone who’s really into this stuff or you’re brand new, there’s going to be something here for you. Now, first on the list is the Deep Seek Harness. This thing’s gotten over 200,000 stars in about 2 weeks, making it the fastest-growing GitHub repo ever. Now, yes, are those
• 0:32 — numbers boosted by a very enthusiastic Chinese base? Sure, but there’s something real here, and it’s because this harness does things that the Claude Code Harness and the Codex Harness just don’t do. And that is it allows you to customize the harness itself at a very base level. Now, what do I mean by that? Well, the Deep Seek Harness mantra is everything is a plugin. So, think of the harness is like the chassis of a car, and we can put whatever we want in there for the engine, we can customize the chairs, the upholstery, everything to an
• 1:04 — extent no one else really can. So, first of all, I can run any model I want with this. Right now, you see the models are coming from Open Router, so literally pretty much every model out there can be used. But the big thing is when it comes to plugins, I can turn and essentially create my own custom plugins that change how the harness works at a very fundamental level. Like I can change how it does tool calling. I can change how it does agent loops. I can change other things that are rather trivial, like I could change the UI and how it actually looks and the colors. Again, things that
• 1:35 — the Claude Code Harnesses doesn’t let you do. Now you can approximate those changes in Claude Code and Codex using things like hooks and even certain skills, but you aren’t changing the fundamental plumbing. So, if you’re someone who’s really into the open-source game, you’re really into tinkering and customizability, Deep Seek Harness is something you should try out. There is a exploding library of different plugins for Deep Seek Harness that change how it works that you can find on GitHub as well. Now number two on the list is Umi Machi and this is something that has started to get a ton of hype over the last couple days. Umi Machi is a custom Linux distribution
• 2:05 — which pretty much combines Arch with AI agents. AI agents become a fundamental part of the actual operating system. A simple example of that is let’s say you have some sort of error with your OS itself. Well, it automatically get sent to your AI agent of choice. You can choose whichever one you want essentially running in the background. That could be Codex, that could be Claude code, it could be Gemini. But the point is it is a fundamental foundational part of this operating system with the Linux package. Now installing and using this is definitely a step above everything else we’re going
• 2:35 — to talk about here. So if you aren’t familiar with Linux, this might be a bridge too far, but if you’re someone who does love Linux and does love AI, I think you should definitely check this out. Now before we hop into the next repo, a quick word from today’s sponsor, me. So I just released the brand new updated Claude code masterclass inside of Chase AI Plus and it is the number one way to go from zero to AI dev, especially if you don’t come from a technical background. We focus on real use cases. I assume you know nothing going into this. So if you’re someone
• 3:05 — who wants to learn how to master the most powerful AI tool out there, this is for you. You can find it linked to it inside of the pin comment. Now third on the list is Any Doc and this comes from Fire crawl, the web scrapers. And what this is is it is a Rust library that converts documents into markdown. So we all know that AI loves markdown, yet tons of us work with your traditional Microsoft Office type documents. Think Word, PowerPoint, Excel, etc. Well, there’s a lot of programs out there that
• 3:35 — allow you to take something that’s Word or PowerPoint and turn it into markdown, but they are either inaccurate or slow and Any Doc kind of solves both of these problems. If we take a look at the benchmarks here and again this is coming from Fire crawl themselves, so take it with a grain of salt, but by the numbers that they’re showing us and they are showing it against some of the big ones like LibreOffice and things like that. It not only crushes all the formats, so it did 14 out of 14 formats versus the others. It does it incredibly fast. The
• 4:05 — median speed was 4.4 milliseconds. Next up was Mammoth, which did it at 52.5. It only did one out of 14 formats. And on the other side Office, which did 12 out of 14 formats, took 1,100 milliseconds. So, you know, it’s faster than everyone. It does more formats and in terms of the score, it also scored better as well. And the way they did the score was they had like a third-party AI judge take a look at it and say, “Well, here’s how complete it
• 4:35 — was. Here’s the structure, etc.” Now again, that’s not to say every single conversion’s going to be perfect, but if you’re someone who has thousands or tens of thousands of documents in these old formats, PowerPoint, Word, whatever, and you want to bring it into markdown, you want AI to be able to better handle it, well, this is perfect for you. Now, number four on the list is Hurder. I love Hurder. I use this every single day and is essentially an upgraded UI for your terminal with a few extra features. So, when I use Hurder, I have my standard terminal like you see right
• 5:05 — here and then on the left, I have some more insights to my agents that are already running and different sort of workspaces. Essentially, it allows me to better multitask when I’m using multiple agents on multiple projects all at the same time. So, if I, you know, spin up an agent here, you see I have agent number one, agent number two. You see I have this workspace, which is my main workspace. I can add a new workspace. Almost think of it as like different sort of folders or different projects I’m working on. Okay, I started up that
• 5:36 — agent. I can see I have workspace one and two, the different agents on and on and on. On top of that, makes it really easy to sort of, you know, split panels, right? I can now have multiple agents in one place. And the really cool thing with this is it is sort of saving your work in a sense. So, if I exit out of this, I can completely exit all those tabs, and then I pull it back up, it’s still going to be there. So, I run Herder again. Boom.
• 6:06 — There is all our stuff. So, it’s just a really lightweight simple addition that is great if you’re someone who uses the terminal and uses multiple agents. Next up, we have Orca sitting at 56,000 stars. Now, if Herder was for those of you who are really into the terminal, then Orca is sort of the equivalent for those of you who are more into VS Code and those sorts of heavier IDEs. That’s because Orca pretty much is VS Code if it was made specifically for AI agents. Some of the standout features include stuff like a mobile companion, so really easy to do remote work. It has a
• 6:37 — built-in design mode. We’re starting to see design modes in a lot of the desktop applications of stuff like Claude, Code, and Codex. But, if you’re someone who doesn’t want to use a desktop app, but still wants a design mode, this is perfect for that. And it also has its own CLI, so you can have AI agents actually control Orca, the IDE itself, instead of just the other way around. It works on macOS, Windows, and Linux. So, if you’re someone who likes the idea of Herder, but you don’t like being purely inside of the terminal, I would go to Orca. Now, number six on the list is ClaudeX Loop, and this is a
• 7:08 — skill for Claude, Code, and Codex. I created the skill myself. I’ve done deep dives on the skill as well. But, the whole point of it is, well, the idea is AI systems do a very bad job of grading their own work. You ask Codex how good Codex’s work was, it’s going to say it was great. You ask Claude how good Claude’s work was, it’s going to say it did awesome. So, how do we solve this? Well, we bring in an outside AI system to grade the work of the first system. So, when Claude creates something or generates a plan, this skill brings in
• 7:38 — Codex, and Codex says, “This is right, this is wrong, this is what we should change. And in fact, this skill essentially gives you an entire four-phase step-by-step process for creating any project or any feature. Starts with reconnaissance, so it goes out on the web, does an actual web search to figure out what it should do, then it’s going to ask you a series of questions. Think of it as an enhanced plan mode before Claude creates a plan, and then Codex and Claude go back and forth adversarially for multiple rounds until they reach a conclusion. Finally,
• 8:09 — one of those systems builds it. It can be either Codex or Claude, and at the end of the build, the second system comes in and says, “Okay, here’s what I think, here’s what I don’t think in terms of what’s right and what’s wrong.” So, the point is you get a second set of eyes the entire process, and in the end, that gives you a better output for less tokens because we aren’t stuck in the scenario where Claude builds something, it looks like it works, it breaks, we go back, we iterate. Like, we just solve it all on the front end. So, if you’re someone who uses Claude Code and Codex, and you want a simple way to combine the
• 8:40 — powers of both, definitely check this one out. The next up, we have Open Montage sitting at just under 54,000 stars, and this is the perfect repo if you’re someone who’s trying to get into AI video production because it’s essentially a giant video production studio in a box. Now, this repo walks you through everything, and there’s a lot of human in the loop involved, so it isn’t as simple as you just giving it a prompt and it executes everything. You have say so as things happen, so you can actually have some sort of creative control. So, it will do research for you, it will do the scripting, it will
• 9:10 — do the prompts, but you’re going to have to provide it some other things like API keys if you want to get the most out of this because if you want good-looking AI video, you do need to use the top models, things like C Dance 2.5. You can’t necessarily rely on very cheap or local video models unless you have a very good computer. But, if you don’t want to pay a cent, there is still stuff you can do with this because you can use some of the graphics that something like Claude Code can create using stuff like Run Motion or Hyper Frames. So, if you’re looking more for like computer-generated type videos, this will also be perfect for you.
• 9:41 — Additionally, it’s set it for specific video pipeline. So, if you want something like an animated explainer versus animation versus a clip factory versus something like a screen demo or a talking head. Essentially, it has different skills for each of these. So, if you want to get into AI video and you’re looking for an all-in-one AI tool to help you achieve that, definitely try out Open Montage. Next on the list, we have Omni route sitting at 58,000 stars and this thing is an AI gateway and it can actually be used completely for free. It gives you essentially a one-stop shop for over 350 AI providers,
• 10:14 — 90 plus which are free and the cool thing is is you can set this up to only use free providers and it will just cycle through all the free providers for you automatically, so you never get like stuck with one task and you used up all your free tokens with that provider and then you need to stop and switch. No, it does it all for you. So, you can just cycle through provider after provider after provider and literally just get free AI. Now, obviously, if it’s a free AI provider, are we necessarily using the frontier models? No, but we can set up this thing so it does use the
• 10:45 — frontier models. I can hook this up to things like Claude code or I can use Codex and I can set up priorities and I can tell it, “Hey, when I want it to switch from one model to the other.” So, it’s turning your computer into a gateway for hundreds of AI models and you can be as sophisticated with the setup as you want or you can keep it as simple as you want it set up. And again, it could be completely free if that’s what you want to do. And so, like I alluded to earlier, it will automatically cycle you through providers. The term that they use here is combos and a combo is a chain of
• 11:15 — models on the route routes across automatically. If something fails, if a quota runs out, it will go to the next healthy model. You don’t have to touch anything and you can set up your own routing strategies. There’s essentially 19 pre-built strategies. So, this is a really cool tool which with as much sophistication as you want it to have. So, if you’re someone who uses, again, a ton of different AI models, or you’re someone who’s looking to essentially token max with free models, you should try this out. And number nine on the list, we have ArchifAI. 30,000 stars, at one point was the number one repo of the
• 11:46 — day, and the sell for ArchifAI is it takes your codebase or any sort of system description and turns it into an artifact, a live, shareable map breaking down how your codebase works. So, you can take a relatively complicated codebase, and you can kind of turn it almost into an Excalidraw type diagram, and you can share this with people, and you can make visual what can be somewhat difficult to understand code, especially if AI wrote everything for you. And this just isn’t a completely static thing that’s just taking snapshots of your
• 12:16 — repo. You can create motion sequences that explain different interactions or different changes you’ve made. You can have it show how data actually flows. So, you can get as complicated with this as you want. So, this is a great repo if you’re someone who, A, is working with complicated codebases, and B, often times has to make some sort of visual explainer to people who maybe aren’t as spun up with how that actually works, because turning it into a visual product like this often times makes that way easier. And last but not least, we have Claude of Tanks. Now, no one’s talked about this game. It only has like 200
• 12:46 — stars, but this is a sick project. This guy essentially recreated World of Tanks in 3.js using just Claude code, and it is wild how cool the graphics are on this, how amazing the physics are and the customization of all of it. It has an in-depth combat system. It has a HUD. It has an X-ray kill cam. It has destructions. Again, the physics are well beyond what you would expect from your typical like Claude built video game. It has multiplayer with local LAN or browser hosting, and I think this is
• 13:17 — cool because it is a perfect example, if you’re someone who wants to create like video game type projects, how to do this using Claude code. Like, it has a bunch of different maps you can choose from. I can choose all these different tanks. I can adjust the camouflage. You can see all the settings on the right-hand side. There’s modules for guns, the engine, the fuel tank, the radio. Like, this is wild. So, shout out to Kevin Lou for creating this and definitely give it a spin. So, that’s where I’m going to end this video. I hope at least a couple of these open-source repos were something
• 13:47 — you haven’t seen before and can integrate into your stack. And as always, make sure to let me know what you thought. Definitely check out Chase A Plus if you want to get your hands on my Cloud Code Masterclass. And I’ll see you around.