10 NEW Github Repos Every Claude User Must Use

CChase AI
Computing/SoftwareVideo & Computer GamesInternet Technology

Transcript

00:00:00So this month, we have seen a ton of brand new,
00:00:02really cool open source repos hit the scene.
00:00:05And in today's video, I'm going to show you 10 of my favorite.
00:00:08And these range from skills you can add to your next cloud code project
00:00:11all the way to completely customizable AI agent harnesses.
00:00:16So whether you're someone who's really into this stuff
00:00:18or you're brand new, there's going to be something here for you.
00:00:22Now, first on the list is the DeepSeek harness.
00:00:25This thing has gotten over 200,000 stars in about two weeks,
00:00:28making it the fastest growing GitHub repo ever.
00:00:31Now, yes, are those numbers boosted by a very enthusiastic Chinese base?
00:00:36Sure, but there's something real here.
00:00:38And it's because this harness does things that the cloud code harness
00:00:41and the codex harness just don't do.
00:00:44And that is it allows you to customize the harness itself at a very base level.
00:00:50Now, what do I mean by that?
00:00:51Well, the DeepSeek harness mantra is everything is a plugin.
00:00:55So think of the harness as like the chassis of a car
00:00:58and we can put whatever we want in there for the engine.
00:01:00We can customize the chairs, the upholstery, everything to an extent.
00:01:05No one else really can.
00:01:06So first of all, I can run any model I want with this.
00:01:10Right now, you see the models are coming from Open Router.
00:01:12So literally pretty much every model out there can be used.
00:01:15But the big thing is when it comes to plugins,
00:01:17I can turn and essentially create my own custom plugins
00:01:21that change how the harness works at a very fundamental level.
00:01:24Like I can change how it does tool calling.
00:01:27I can change how it does agent loops.
00:01:29I can change other things that are rather trivial.
00:01:31Like I could change the UI and how it actually looks and the colors.
00:01:34Again, things that the cloud code harness just doesn't let you do.
00:01:37Now, you can approximate those changes in cloud code and codex
00:01:39using things like hooks and even certain skills,
00:01:41but you aren't changing the fundamental plumbing.
00:01:44So if you're someone who's really into the open source game,
00:01:46you're really into tinkering and customizability,
00:01:48DeepSeek Harness is something you should try out.
00:01:50There is a exploding library of different plugins
00:01:54for DeepSeek Harness that change how it works
00:01:56that you can find on GitHub as well.
00:01:58Now, number two on the list is Umachi.
00:02:00And this is something that has started to get a ton of hype
00:02:02over the last couple of days.
00:02:03So Umachi is a custom Linux distribution,
00:02:05which pretty much combines Arch with AI agents.
00:02:08AI agents become a fundamental part of the actual operating system.
00:02:12A simple example of that is,
00:02:13let's say you have some sort of error with your OS itself.
00:02:16Well, it automatically gets sent to your AI agent of choice.
00:02:19You can choose whichever one you want,
00:02:20essentially running in the background.
00:02:22That could be codex, that could be cloud code, it could be Gemini.
00:02:24But the point is, it is a fundamental foundational part
00:02:28of this operating system with the Linux package.
00:02:32Now, installing and using this
00:02:33is definitely a step above everything else we're going to talk about here.
00:02:36So if you aren't familiar with Linux, this might be a bridge too far.
00:02:39But if you're someone who does love Linux and does love AI,
00:02:43I think you should definitely check this out.
00:02:45Now, before we hop into the next repo,
00:02:46a quick word from today's sponsor, me.
00:02:48So I just released the brand new updated
00:02:51Cloud Code Masterclass inside of Chase AI Plus.
00:02:54And it is the number one way to go from zero to AI dev,
00:02:57especially if you don't come from a technical background.
00:03:00We focus on real use cases.
00:03:01I assume you know nothing going into this.
00:03:04So if you're someone who wants to learn how to master
00:03:06the most powerful AI tool out there, this is for you.
00:03:10You can find a link to it inside of the pinned comment.
00:03:13Now, third on the list is AnyDoc.
00:03:15And this comes from Firecrawl, the web scrapers.
00:03:18And what this is, is it is a Rust library
00:03:21that converts documents into Markdown.
00:03:24So we all know that AI loves Markdown,
00:03:27and yet tons of us work with your traditional
00:03:29Microsoft Office type documents.
00:03:31Think Word, PowerPoint, Excel, et cetera.
00:03:33Well, there's a lot of programs out there
00:03:34that allow you to take something that's Word or PowerPoint
00:03:37and turn it into Markdown,
00:03:38but they are either inaccurate or slow.
00:03:41And AnyDoc kind of solves both of these problems.
00:03:45If we take a look at the benchmarks here,
00:03:48and again, this is coming from Firecrawl themselves,
00:03:50so take it with a grain of salt.
00:03:51But by the numbers that they're showing us,
00:03:54and they are showing it against some of the big ones
00:03:56like LibreOffice and things like that,
00:03:58it not only crushes all the formats,
00:04:00so it did 14 out of 14 formats versus the others,
00:04:03it does it incredibly fast.
00:04:05The median speed was 4.4 milliseconds.
00:04:09Next up was Mammoth, which did it at 52.5.
00:04:12It only did one out of 14 formats.
00:04:15And on the other side,
00:04:17Office, which did 12 out of 14 formats,
00:04:19took 1100 milliseconds.
00:04:22So, you know, it's faster than everyone.
00:04:25It does more formats.
00:04:27And in terms of the score,
00:04:28it also scored better as well.
00:04:30And the way they did the score
00:04:31was they had like a third-party AI judge
00:04:33take a look at it and say,
00:04:34well, here's how complete it was.
00:04:35Here's the structure, et cetera.
00:04:37Now, again, that's not to say
00:04:37every single conversion is going to be perfect,
00:04:40but if you're someone who has thousands
00:04:42or tens of thousands of documents
00:04:44in these old formats,
00:04:46PowerPoint, Word, whatever,
00:04:47and you want to bring it into Markdown,
00:04:48you want AI to be able to better handle it,
00:04:50this is perfect for you.
00:04:52Now, number four on the list is Herder.
00:04:54I love Herder.
00:04:55I use this every single day
00:04:56and is essentially an upgraded UI
00:04:58for your terminal with a few extra features.
00:05:01So when I use Herder,
00:05:03I have my standard terminal
00:05:04like you see right here.
00:05:06And then on the left,
00:05:07I have some more insights
00:05:09to my agents that are already running
00:05:11and different sort of workspaces.
00:05:13Essentially, it allows me
00:05:14to better multitask
00:05:16when I'm using multiple agents
00:05:18on multiple projects
00:05:19all at the same time.
00:05:20So if I, you know,
00:05:22spin up an agent here,
00:05:24you see I have agent number one,
00:05:26agent number two.
00:05:27You see I have this workspace,
00:05:28which is my main workspace.
00:05:29I can add a new workspace.
00:05:30Almost think of it
00:05:31as like different sort of folders
00:05:33or different projects I'm working on.
00:05:35Okay, I started up that agent.
00:05:36I can see I have workspace one and two,
00:05:38the different ages,
00:05:39on and on and on.
00:05:40On top of that,
00:05:41it makes it really easy
00:05:42to sort of, you know,
00:05:43split panels, right?
00:05:45I can now have multiple agents
00:05:46in one place.
00:05:47And the really cool thing with this
00:05:49is it is sort of saving your work
00:05:51in a sense.
00:05:53So if I exit out of this,
00:05:55I completely exit all those tabs
00:05:57and then I pull it back up,
00:05:59it's still going to be there.
00:06:03So I run Herder again.
00:06:05Boom.
00:06:06There is all our stuff.
00:06:08So it's just a really lightweight,
00:06:10simple addition that is great
00:06:12if you're someone who uses the terminal
00:06:13and uses multiple agents.
00:06:16Next up, we have Orca
00:06:17sitting at 56,000 stars.
00:06:19Now, if Herder was for those of you
00:06:21who are really into the terminal,
00:06:22then Orca is sort of the equivalent
00:06:24for those of you who are more into VS Code
00:06:27and those sorts of heavier IDEs.
00:06:29That's because Orca pretty much is VS Code
00:06:30if it was made specifically for AI agents.
00:06:33Some of the standout features
00:06:34include stuff like a mobile companion,
00:06:35so really easy to do remote work.
00:06:37It has a built-in design mode.
00:06:39We're starting to see design modes
00:06:40in a lot of the desktop applications
00:06:43of stuff like Claude Code and Codex,
00:06:45but if you're someone
00:06:46who doesn't want to use a desktop app
00:06:47but still wants a design mode,
00:06:48this is perfect for that.
00:06:49And it also has its own CLI,
00:06:51so you can have AI agents
00:06:52actually control Orca, the IDE itself,
00:06:56instead of just the other way around.
00:06:57It works on macOS, Windows, and Linux.
00:06:59So if you're someone
00:07:00who likes the idea of Herder,
00:07:01but you don't like being purely
00:07:03inside of the terminal,
00:07:04I would go to Orca.
00:07:05Now, number six on the list
00:07:06is Claudex Loop,
00:07:07and this is a skill
00:07:08for Claude Code and Codex.
00:07:10I created this skill myself.
00:07:12I've done deep dives
00:07:13on this skill as well,
00:07:14but the whole point of it is,
00:07:16well, the idea is AI systems
00:07:18do a very bad job
00:07:19of grading their own work.
00:07:20You ask Codex
00:07:21how good Codex work was,
00:07:22it's going to say it was great.
00:07:23You ask Claude
00:07:24how good Claude's work was,
00:07:25it's going to say it did awesome.
00:07:27So how do we solve this?
00:07:28Well, we bring in an outside AI system
00:07:30to grade the work
00:07:31of the first system.
00:07:33So when Claude creates something
00:07:35or generates a plan,
00:07:37this skill brings in Codex
00:07:38and Codex says,
00:07:39this is right, this is wrong,
00:07:40this is what we should change.
00:07:42And in fact,
00:07:42this skill essentially gives you
00:07:45an entire four-phase
00:07:47step-by-step process
00:07:48for creating any project
00:07:49or any feature.
00:07:51Starts with reconnaissance,
00:07:52so it goes out on the web,
00:07:53does an actual web search
00:07:54to figure out what it should do.
00:07:56Then it's going to ask you
00:07:57a series of questions,
00:07:58think of it as an enhanced plan mode
00:08:00before Claude creates a plan
00:08:02and then Codex and Claude
00:08:04go back and forth adversarially
00:08:05for multiple rounds
00:08:06until they reach a conclusion.
00:08:08Finally,
00:08:09one of those systems builds it.
00:08:10It can be either Codex or Claude
00:08:11and at the end of the build,
00:08:13the second system comes in
00:08:14and says,
00:08:15okay, here's what I think,
00:08:16here's what I don't think
00:08:17in terms of what's right
00:08:18and what's wrong.
00:08:19So the point is
00:08:20you get a second set of eyes,
00:08:22the entire process
00:08:23and in the end,
00:08:24that gives you a better output
00:08:26for less tokens
00:08:27because we aren't stuck
00:08:28into the scenario
00:08:29where Claude builds something,
00:08:30it looks like it works,
00:08:31it breaks,
00:08:31we go back,
00:08:32we iterate,
00:08:33like we just solve it
00:08:34all on the front end.
00:08:35So if you're someone
00:08:36who uses Claude code
00:08:37and Codex
00:08:37and you want a simple way
00:08:39to combine the powers of both,
00:08:41definitely check this one out.
00:08:42Now next up,
00:08:42we have Open Montage
00:08:44sitting at just under 54,000 stars
00:08:46and this is the perfect repo
00:08:47if you're someone
00:08:48who's trying to get into
00:08:49AI video production
00:08:51because it's essentially
00:08:52a giant video production studio
00:08:54in a box.
00:08:56Now this repo walks you
00:08:57through everything
00:08:58and there's a lot of human
00:08:59in the loop involved
00:09:00so it isn't as simple
00:09:00as you just giving it a prompt
00:09:01and it executes everything.
00:09:03You have say so
00:09:04as things happen
00:09:05so you can actually have
00:09:06some sort of creative control.
00:09:07So it will do research for you,
00:09:08it will do the scripting,
00:09:10it will do the prompts
00:09:11but you're going to have
00:09:12to provide it some other things
00:09:13like API keys
00:09:14if you want to get the most
00:09:15out of this
00:09:15because if you want
00:09:16good looking AI video,
00:09:17you do need to use
00:09:18the top models
00:09:19things like Seed Dance 2.5
00:09:21you can't necessarily rely
00:09:22on very cheap
00:09:23or local video models
00:09:24unless you have
00:09:25a very good computer
00:09:26but if you don't want
00:09:27to pay a cent
00:09:28there is still stuff
00:09:28you can do with this
00:09:29because you can use
00:09:30some of the graphics
00:09:31that something like
00:09:31Claude Code can create
00:09:32using stuff like
00:09:33Remotion or Hyper Frames
00:09:35so if you're looking
00:09:36more for like computer
00:09:37generated type videos
00:09:39this will also be
00:09:40perfect for you.
00:09:41Additionally,
00:09:41it's set up
00:09:42for specific video pipelines
00:09:43so if you want
00:09:44something like
00:09:45an animated explainer
00:09:46versus animation
00:09:47versus a clip factory
00:09:48versus something
00:09:49like a screen demo
00:09:49or a talking head
00:09:50it essentially
00:09:51has different skills
00:09:51for each of these
00:09:52so if you want
00:09:53to get into AI video
00:09:54and you're looking
00:09:55for an all-in-one AI tool
00:09:56to help you achieve that
00:09:57definitely try out
00:09:58Open Montage.
00:09:59Next on the list
00:10:00we have OmniRoute
00:10:01sitting at 58,000 stars
00:10:03and this thing
00:10:03is an AI gateway
00:10:05and it can actually
00:10:06be used completely
00:10:07for free.
00:10:08It gives you
00:10:09essentially a one-stop shop
00:10:11for over 350 AI providers
00:10:1390 plus which are free
00:10:15and the cool thing is
00:10:17is you can set this up
00:10:18to only use free providers
00:10:19and it will just cycle
00:10:21through all the free providers
00:10:23for you automatically
00:10:24so you never get like
00:10:25stuck with one task
00:10:27and you used up
00:10:28all your free tokens
00:10:28with that provider
00:10:29and then you need
00:10:30to stop and switch.
00:10:30No, it does it all for you
00:10:31so you can just cycle
00:10:33through provider
00:10:34after provider
00:10:34after provider
00:10:35and literally
00:10:36just get free AI.
00:10:37Now, obviously
00:10:39if it's a free AI provider
00:10:40are we necessarily
00:10:41using the frontier models?
00:10:42No, but we can set up
00:10:44this thing so it does
00:10:45use the frontier models.
00:10:46I can hook this up
00:10:47to things like Claude Code
00:10:48or I can use Codex
00:10:49and I can set up priorities
00:10:51and I can tell it
00:10:52hey, when I want it
00:10:53to switch from one model
00:10:54to the other
00:10:54so it's turning
00:10:55your computer
00:10:56into a gateway
00:10:57for hundreds of AI models
00:10:59and you can be
00:11:00as sophisticated
00:11:00with the setup
00:11:01as you want
00:11:01or you can keep it
00:11:02as simple
00:11:03as you want it set up
00:11:04and again
00:11:05it could be completely free
00:11:06if that's what you want to do
00:11:07and so like I alluded to earlier
00:11:09it will automatically cycle
00:11:10you through providers.
00:11:11a term that they use here
00:11:12is combos
00:11:13and a combo
00:11:14is a chain of models
00:11:15on the route
00:11:16routes across automatically.
00:11:18If something fails
00:11:19if a quarter runs out
00:11:20it will go to the next
00:11:21healthy model
00:11:22you don't have to touch anything
00:11:23and you can set up
00:11:24your own routing strategies
00:11:25there's essentially
00:11:2619 pre-built strategies
00:11:28so this is a really cool tool
00:11:30which with as much sophistication
00:11:32as you want it to have
00:11:33so if you're someone
00:11:34who uses
00:11:35again a ton of different
00:11:36AI models
00:11:36or you're someone
00:11:37who's looking to
00:11:38essentially token max
00:11:39with free models
00:11:40you should try this out.
00:11:41The number nine
00:11:42on the list
00:11:42we have Archify
00:11:4330,000 stars
00:11:44at one point
00:11:45was the number one
00:11:45repo of the day
00:11:47and the sell for Archify
00:11:48is it takes your code base
00:11:50or any sort of system description
00:11:51and turns it into
00:11:53an artifact
00:11:54a live shareable map
00:11:56breaking down
00:11:57how your code base works
00:11:58so you can take
00:11:59a relatively complicated code base
00:12:00and you can kind of turn it
00:12:02almost into an Excalibur
00:12:03type diagram
00:12:04and you can share this
00:12:05with people
00:12:06and you can make visual
00:12:07what can be somewhat difficult
00:12:09to understand code
00:12:10especially if AI
00:12:11wrote everything for you
00:12:12and this just isn't
00:12:13a completely static thing
00:12:14that's just taking
00:12:15snapshots of your repo
00:12:16you can create motion sequences
00:12:18that explain
00:12:19different interactions
00:12:20or different changes
00:12:21you've made
00:12:21you can have it show
00:12:22how data actually flows
00:12:23so you can get as complicated
00:12:24with this as you want
00:12:26so this is a great repo
00:12:27if you're someone who
00:12:27A. is working with
00:12:28complicated code bases
00:12:29and B. oftentimes
00:12:31has to make some sort of
00:12:32visual explainer
00:12:33two people
00:12:33who maybe aren't
00:12:34as spun up
00:12:35with how that actually works
00:12:37because turning it
00:12:38into a visual product
00:12:39like this
00:12:39oftentimes makes that
00:12:40way easier
00:12:41and last but not least
00:12:42we have Claude of Tanks
00:12:44now no one's talked
00:12:45about this game
00:12:45it only has like 200 stars
00:12:47but this is a sick project
00:12:48this guy essentially
00:12:49recreated World of Tanks
00:12:51in 3.js
00:12:52using just Claude code
00:12:54and it is wild
00:12:56how cool the graphics
00:12:57are on this
00:12:58how amazing
00:12:59the physics are
00:13:00and the customization
00:13:01of all of it
00:13:02it has an in-depth
00:13:03combat system
00:13:04it has a HUD
00:13:04it has an x-ray
00:13:05kill cam
00:13:06it has destructions
00:13:07again the physics
00:13:08are well beyond
00:13:09what you would expect
00:13:10from your typical
00:13:11like Claude build video game
00:13:13it has multiplayer
00:13:14with local LAN
00:13:15or browser hosting
00:13:16and I think this is cool
00:13:17because it is a perfect example
00:13:19if you're someone
00:13:20who wants to create
00:13:21like video game type projects
00:13:22how to do this
00:13:23using Claude code
00:13:24like it has a bunch
00:13:26of different maps
00:13:26you can choose from
00:13:28I can choose
00:13:28all these different tanks
00:13:29I can adjust
00:13:30the camouflage
00:13:31you can see
00:13:32all the settings
00:13:33on the right hand side
00:13:34there's modules
00:13:34for guns
00:13:35the engine
00:13:35the fuel tank
00:13:36the radio
00:13:36like this is wild
00:13:38so shout out
00:13:39to Kevin Liu
00:13:40for creating this
00:13:41and definitely
00:13:41give it a spin
00:13:42so that's where
00:13:43I'm going to end this video
00:13:43I hope at least
00:13:44a couple of these
00:13:45open source repos
00:13:46were something
00:13:47you haven't seen before
00:13:48and can integrate
00:13:48into your stack
00:13:49and as always
00:13:50make sure to let me know
00:13:51what you thought
00:13:52definitely check out
00:13:53Chase AI Plus
00:13:54if you want to get your hands on my Cloud Code Masterclass
00:13:56and I'll see you around.

Key Takeaway

Ten open-source repositories enhance AI workflows, featuring the DeepSeek harness for base-level customization, AnyDoc for rapid Markdown conversion, and OmniRoute for automated free token routing.

Highlights

  • The DeepSeek harness achieved over 200,000 stars in two weeks, establishing it as the fastest-growing GitHub repository.

  • AnyDoc converts documents into Markdown across 14 formats with a median speed of 4.4 milliseconds.

  • OmniRoute functions as an AI gateway connecting over 350 AI providers, including more than 90 free options.

  • Claudex Loop utilizes an adversarial dual-system process where Codex evaluates and refines Claude Code outputs.

  • Claude of Tanks implements a fully functional 3.js multiplayer game featuring an in-built x-ray kill cam and physics engine, built entirely through Claude Code.

Timeline

DeepSeek Harness and Umachi OS

  • The DeepSeek harness operates on an everything-is-a-plugin mantra that allows deep customizability of agent loops and tool calling.
  • Umachi integrates Arch Linux with background AI agents to automatically handle operating system errors.

Open-source developer tools have shifted toward deep modifiability. The DeepSeek harness provides chassis-level control over AI execution pipelines, letting developers swap models and rewrite core behaviors. Meanwhile, Umachi embeds AI agents directly into the Linux package ecosystem to monitor and resolve operating system faults automatically.

AnyDoc Document Conversion Library

  • AnyDoc is a Rust-based library that converts legacy Office documents into AI-ready Markdown.
  • Benchmarks show a median processing speed of 4.4 milliseconds across 14 supported formats.

Legacy file formats like Word, Excel, and PowerPoint create friction for AI pipelines that rely on Markdown. AnyDoc addresses this bottleneck by processing document conversions significantly faster and more accurately than traditional tools like LibreOffice or Mammoth.

Herder Terminal UI and Orca IDE

  • Herder provides a multi-workspace terminal interface that persists running agent tabs across sessions.
  • Orca adapts the VS Code environment specifically for AI agents with a mobile companion and built-in design mode.

Managing multiple concurrent AI agents requires specialized interface layers. Herder organizes terminal workflows into distinct workspaces that retain state upon exit. For developers preferring graphical environments, Orca integrates agent controls directly into a heavier IDE structure.

Claudex Loop and Open Montage Studio

  • Claudex Loop uses an adversarial four-phase process where Codex and Claude evaluate each other's work.
  • Open Montage offers a human-in-the-loop studio workflow for automated AI video production pipelines.

Single AI systems struggle to objectively grade their own code generation. Claudex Loop solves this by introducing multi-model opposition to catch errors early. In parallel, Open Montage structures end-to-end video production workflows across scripting, prompting, and visual rendering.

OmniRoute, Archify, and Claude of Tanks

  • OmniRoute automatically cycles across more than 90 free AI providers and 350 total models via pre-built routing strategies.
  • Archify transforms complex codebases into live shareable visual maps and data flow sequences.
  • Claude of Tanks demonstrates game development potential by running a 3.js multiplayer tank battle engine built entirely through Claude Code.

Token optimization and visualization tools maximize developer efficiency. OmniRoute acts as an intelligent gateway that bypasses depleted quotas by rotating through free model endpoints. Archify generates architectural diagrams from raw code, while experimental projects like Claude of Tanks push the boundaries of AI-generated applications.

Community Posts

View all posts