10 NEW Github Repos Every Claude User Must Use

At a glance

Chase Hannegan counts down ten open source repositories that surfaced in the last month or so, from an agent framework that picked up over 200,000 GitHub stars in about two weeks to a fan built recreation of World of Tanks sitting at roughly 200. Every entry connects back to Claude Code or Codex in some way: some replace the terminal or IDE you run them in, some feed them documents or video, one makes the two agents police each other's work, and one is simply proof of what they can build with nobody touching the keyboard. Chase demonstrates each tool live on his own machine, states plainly who should install it and who should skip it, and closes with a plug for his own paid Claude Code course. This page treats each of the ten as a standalone reference: what it is, the problem it solves, what Chase actually shows and claims, and what it takes to go try it yourself.

The ten, in the order Chase ranks them

Chase opens with no throat clearing: a bunch of interesting open source repositories turned up over the past month, and he is running through ten favorites, from skills you drop into an existing Claude Code project to fully customizable tooling for AI agents. He tells beginners and long time users alike that there is something in the list for them, then goes straight into the countdown.

Read top to bottom, the ten sort into a rough stack: an operating system at the base, an agent framework that runs on top of it, a gateway that decides which model answers, two different front ends for actually watching the agent work, a skill that keeps two agents honest with each other, two tools that feed the agent raw material (documents in, video out), a tool that turns the agent's own output into something a human can follow, and, right at the top, a finished game that exists only because all of the above eventually works.

1. DeepSeek Harness: rebuilding the agent's plumbing, not just prompting it

First on Chase's list is DeepSeek Harness, which he says gathered over 200,000 stars in about two weeks, making it, in his words, the fastest growing GitHub repository in history. He does not pretend that number is clean: "are these numbers backed up by a very enthusiastic Chinese community? Of course," he says, but he argues there is something real underneath the hype, because the tool does something neither Claude Code nor Codex lets you do, which is customize the framework itself at a fundamental level.

The mantra of the project, and its own tagline on GitHub, is that everything is a plugin. Chase's analogy: think of the harness as a car chassis. You can put any engine in it, customize the seats and the upholstery, to a degree no competing tool allows. Concretely, that means the model layer runs through OpenRouter, so literally any existing model can sit behind it, and beyond that you can write custom extensions that change how tools are invoked, how agent workflows run, or even small things like the interface's look and colors. He is explicit that you can approximate some of this in Claude Code or Codex using hooks or skills, but you cannot touch the actual "plumbing" of the system the way you can here. There is also a growing library of community plugins for the harness already on GitHub.

His verdict: if you are an open source fan who likes to experiment and wants the ability to customize, try DeepSeek Harness. If you just want to get work done without touching the engine, it is not built for you.

To actually run it, the project is a developer preview from DeepSeek built on a plugin runtime called Cordis, where models, tools, skills, sessions, sandboxes and even scheduling are all separately swappable plugins mounted onto the core. It ships breaking changes as it iterates, so Chase's "if you like to experiment" framing is also a fair warning.

2. Omarchy: an operating system with an agent wired into it

Second is Omarchy, a custom Linux distribution that combines Arch Linux with AI agents. The pitch, in Chase's phrasing, is that AI agents are becoming part of the operating system itself, not an app you launch inside it. His example: if something throws an error at the OS level, that error gets automatically sent to your AI agent of choice, running in the background, whether that is Codex, Claude Code or Gemini. The agent ships as part of the operating system, distributed as a Linux package rather than bolted on afterward.

Chase is upfront that this is a bigger commitment than everything else on the list: "installing and using this is definitely a level above anything else we'll talk about here." If you are not already comfortable with Linux, he thinks it will be too complicated. If you already love both Linux and AI, he says give it a try.

What the video does not say, because it predates this episode, is whose project this is: Omarchy is built by DHH, the creator of Ruby on Rails, and pairs Arch with the Hyprland tiling window manager and a custom shell. By the time of this video it ships ten coding agent CLIs pre wired as launchers, Claude Code and Codex among them, which is exactly the "agents baked into the OS" behavior Chase is demonstrating.

Right after Omarchy, Chase breaks for a sponsor read, and the sponsor is himself: a plug for the newly updated Claude Code masterclass inside Chase AI Plus, aimed at people with no technical background, with a link in the pinned comment. He returns to the list immediately after.

3. Anydoc: fourteen document formats to Markdown in single digit milliseconds

Third is Anydoc, built by Firecrawl, the web scraping company. It is a Rust library that converts documents to Markdown. The problem it targets: AI systems work best on Markdown, but most real world documents are Word, PowerPoint or Excel files, and the existing converters for those formats are either inaccurate or slow. Chase presents Firecrawl's own published benchmark, with the caveat, said out loud, to "take it with a grain of salt" since it is the vendor's own data. By his numbers, Anydoc handles all 14 of 14 tested formats with a median conversion time of 4.4 milliseconds. Mammoth, a comparison tool, handles only 1 of the 14 formats in 52.5 milliseconds. LibreOffice, one of the "giants" he says Anydoc beats, handles 12 of the 14 formats but takes about 1,100 milliseconds, roughly 250 times slower. Quality was judged by a third party AI model comparing outputs for completeness and structure; Chase is careful to note that beating the field on speed and coverage does not mean every single conversion comes out perfect.

Worth knowing beyond what Chase says on screen: Firecrawl's own published benchmark (482 judged comparisons, scored by Claude Sonnet 5 against a page rendered ground truth) puts Anydoc's quality score at 81 out of 100 against LibreOffice's 87, so the tradeoff is real, Anydoc is dramatically faster and covers more formats, LibreOffice is a little more faithful to the original when it can handle the format at all. For anyone converting thousands of legacy Office documents so an LLM can read them, that tradeoff is normally an easy one to accept.

His verdict: if you have a pile of old Word or PowerPoint files to turn into something an AI system can ingest, Anydoc is the tool built for exactly that.

4. Herdr: a terminal that remembers every agent you were running

Fourth is Herdr, which Chase says he uses every day: an improved terminal interface with a sidebar of extra information about whatever agents and workspaces are currently running. On screen he shows a normal terminal with a panel on the left listing agent one and agent two, plus a "main space" workspace that he can duplicate into a new one, each workspace standing in for a separate folder or project. He can split panels to hold multiple agents in one place at once.

The feature he singles out as the coolest: it saves your state. He closes every tab, quits the app, reopens it, and everything, agents, workspaces and layout, is exactly where he left it. He calls it a lightweight, simple add on, ideal for anyone who lives in a terminal and runs multiple agents at once.

To actually use it, Herdr is a single Rust binary with no external dependencies, built as a client server terminal multiplexer where the server owns the running processes so a detach never kills your agents. It auto detects more than 15 coding agent CLIs, Claude Code and Codex included, with zero configuration.

5. Orca: the IDE version of the same idea, with a phone in your pocket

Fifth is Orca, which Chase frames directly against Herdr: "if Herdr was for those who are passionate about the terminal, Orca is a kind of equivalent for those who prefer VS Code and other more complex development environments." It has 56,000 stars by his count, and it is effectively VS Code rebuilt from scratch for AI agents. He calls out a mobile companion app that makes remote supervision much easier, and a built in design mode, noting that design modes are starting to show up across many desktop AI tools, including Claude Code and Codex themselves, so an IDE with the same feature but a full desktop app around it fills a gap. Orca also ships its own CLI, which flips the usual direction: AI agents can control Orca the IDE, instead of only the other way around. It runs on macOS, Windows and Linux.

His verdict: if you like what Herdr does but do not want to work exclusively inside a terminal, choose Orca.

Beyond what Chase demonstrates, Orca is built to run several agent CLIs, Claude Code, Codex, Cursor, OpenCode or Pi, side by side, each isolated in its own Git worktree, and its design mode works by letting you click any element in an embedded Chromium browser and injecting the resulting HTML, CSS and a screenshot straight into the agent's prompt.

6. Claudex-Loop: making Claude and Codex grade each other

Sixth is Claudex-Loop, a Claude Code and Codex skill that Chase built himself and has separately reviewed in detail on his channel. The problem it exists to solve, in his words: AI systems are very bad at evaluating their own work. "Ask Codex how well he did the job, and he'll say great. Ask Claude, and he'll say he did an impressive job." His fix is to bring in a third party AI system to grade the first one. When Claude produces something, or generates a plan, the skill hands it to Codex, which reports back what is right, what is wrong, and what needs to change.

The skill runs a full four phase process for building any project or feature. It starts with exploration, where the agent searches online to work out what to do, then moves into a round of questions, functioning as an advanced planning mode before Claude drafts a plan. From there, Codex and Claude go back and forth, competing over several rounds until they land on a shared conclusion. Finally, one of the two, Codex or Claude, actually builds the project, and once the build is done, the other system steps back in to say what it thinks is right and wrong about the result. The upshot, per Chase, is a second pair of eyes across the entire process and a better outcome using fewer tokens overall, because you skip the usual loop of building something that looks fine, watching it break later, and going back to fix it. The decisions get made up front instead.

His verdict: if you already use both Claude Code and Codex and want to combine what each is good at, this is worth a look.

7. OpenMontage: a video studio in a box, with you still holding the wheel

Seventh is OpenMontage, sitting just under 54,000 stars, which Chase describes as a huge video studio in a box for anyone trying to get into AI video production. He is clear that it is not a one shot, hands off tool: there is real human work in the loop, so you keep creative control rather than handing everything to the agent. It will research a topic for you, write a script, and prepare prompts, but you still have to supply things like API keys if you want the best possible output, because top tier AI video quality currently requires paid, frontier models rather than cheap or local ones, unless you own a very powerful machine. The model he names on screen is Seedance 2.5, ByteDance's current flagship video model. If you would rather not spend anything, he says you can still produce something using visuals a coding agent builds directly, naming Remotion and HyperFrames as the free route, both of which render programmatic, code driven video rather than generating pixels from a prompt.

OpenMontage is also tuned for specific pipelines rather than one generic mode: he lists animated explainer videos, general animation, a "clip factory," and screen demo or talking head style videos, and says the project ships different skills for each of those jobs.

His verdict: if you want to get into AI video and want a tool flexible enough to cover more than one style of output, try OpenMontage.

8. OmniRoute: one gateway to 350+ AI providers, 90 of them free

Eighth is OmniRoute, at 58,000 stars by Chase's count, which he calls a completely free AI gateway, a one stop shop for over 350 AI providers, 90 of which cost nothing. The feature he leans on hardest: you can set it to use only free providers, and it will automatically switch between them for you, so you are never stuck burning through one provider's free allowance and then having to manually hunt for the next. He is honest that free does not automatically mean cutting edge, but says you can configure the system toward more advanced models, wire it into Claude Code or Codex, set your own priorities, and decide exactly when it should switch. The project's own term for this is a "combo," a chain of models along one route; routing happens automatically, and if something fails or a quota runs out, OmniRoute moves to the next working model without you touching anything. There are 19 ready made routing strategies out of the box, and you can build custom ones. Chase's summary: it turns your computer into a gateway to hundreds of AI models, and you decide how complicated to make it, all the way down to entirely free.

His verdict: if you regularly use many different AI models, or want to squeeze the most out of free token allowances, this is worth trying.

9. Archify: turning a codebase AI wrote into something a human can read

Ninth is a project Chase's on screen chapter title calls "Archy," at 30,000 stars and, for a while, the number one trending repository on GitHub for the day; the project's actual published name is Archify, which is what you will find if you search for it directly. Its core trick is turning your codebase, or any system description, into what Chase calls an artifact: an interactive, shareable map that explains how the code actually works. He describes taking a fairly complex codebase and turning it into something like an Excalidraw style diagram that you can hand to someone else, which matters especially for code written entirely by an AI agent, where the person reading it did not write a line of it.

He is careful to note it is not just a static snapshot tool: you can build sequences of movement that explain specific interactions or changes over time, and customize how data flows through the diagram, so it can be as simple or as complicated as the underlying system demands.

His verdict: if you work on codebases and regularly have to explain how they function to people who cannot just read the code, turning it into a visual product like this makes that job much easier.

10. Claude of Tanks: proof, not a tool

Last on the list is Claude of Tanks, which Chase says nobody else has covered yet, and which sits at only around 200 stars. He calls it an incredible project anyway: a recreation of World of Tanks built in three.js, made entirely with Claude Code. He praises the graphics, calls the physics "far superior to what you would expect from a typical Claude built video game," and highlights a deep combat system, a full interface, and an X-ray camera that shows hits and destruction in detail. The game supports multiplayer over LAN or browser hosting. On screen he flips through a set of different maps, a roster of selectable tanks, customizable camouflage, and a settings panel covering individual modules, guns, engine, fuel tank and radio. He thanks its creator, whom he credits on screen as Kevin Lu; the repository itself credits Kevin B. Liu as the person who designed, directed and wrote every line of original gameplay code, with Claude and Codex assisting on research, vehicle modeling, simulation, networking, performance and deployment.

His framing: a cool example for anyone who wants to build a video game using Claude Code, and worth going to actually play.

Chase closes the video there, hoping the viewer discovered at least a few repositories they had not seen before, inviting comments on what people think, and pointing once more to Chase AI Plus and his Claude Code masterclass.

Ten at a glance

Key takeaways

• Every one of Chase's ten picks either extends what Claude Code and Codex can do, changes where you sit while running them, or exists as a finished demonstration of what they can build.
• The commitment level varies enormously, from a Claude Code skill you add in minutes (Claudex-Loop) to swapping your entire operating system (Omarchy). Chase is explicit about which end of that spectrum each pick sits on, and who should skip the heavier ones.
• Big star counts are not proof of quality on their own. Chase says so directly about DeepSeek Harness's community boosted growth, and the same caution applies to any vendor supplied benchmark, including Firecrawl's own numbers for Anydoc.
• A recurring theme across four of the ten, DeepSeek Harness, Herdr and Orca, Claudex-Loop, and OmniRoute, is that a single agent, a single terminal, or a single model provider becomes a bottleneck, and each tool removes a different one of those bottlenecks.
• Claude of Tanks is the outlier: not a tool for building with Claude Code, but evidence of what a disciplined agent pipeline can ship when a human still directs the design.

Chapters

• 0:00 Intro
• 0:22 DeepSeek Harness
• 1:58 Omarchy
• 3:13 Anydoc
• 4:52 Herdr
• 6:15 Orca
• 7:05 Claudex-Loop
• 8:43 OpenMontage
• 9:59 OmniRoute
• 11:41 Archy
• 12:41 Claude of Tanks

Notable quotes

The mantra of the DeepSeek harness is that everything is a plugin. Chase Hannegan, on DeepSeek Harness, 0:35

Are these numbers backed up by a very enthusiastic Chinese community? Of course. But there's something real here. Chase Hannegan, on DeepSeek Harness's star count, 0:30

AI agents are becoming a part of the operating system itself. Chase Hannegan, on Omarchy, 2:00

Ask Codex how well he did the job, and he'll say great. Ask Claude, and he'll say he did an impressive job. Chase Hannegan, on why Claudex-Loop exists, 7:10

This turns your computer into a gateway to hundreds of AI models. Chase Hannegan, on OmniRoute, 10:40

The physics are far superior to what you would expect from a typical Claude built video game. Chase Hannegan, on Claude of Tanks, 13:10

Resources mentioned

• DeepSeek Harness by DeepSeek, and its model routing layer OpenRouter
• Omarchy (omarchy.org), built on Arch Linux and Hyprland by DHH
• Anydoc by Firecrawl, benchmarked against Mammoth and LibreOffice
• Herdr
• Orca, compared by Chase to VS Code
• Claudex-Loop, Chase's own Claude Code and Codex skill
• OpenMontage, with Seedance 2.5 as the paid model named on screen and Remotion plus HyperFrames as the free route
• OmniRoute
• Archify, shown on screen as "Archy," compared by Chase to Excalidraw
• Claude of Tanks (play it at claudeoftanks.kevinliu.studio), built in three.js by Kevin B. Liu
• Claude Code and Anthropic, and Codex from OpenAI, the two agents every repository on this list orbits
• Chase AI's channel and his paid course, Chase AI Plus

Where this leaves you

Nothing here is presented by Chase as a verdict on which single repository to install; the video works as a survey, and the honest way to read it is by commitment level rather than by star count. Claudex-Loop, Herdr and OmniRoute are close to zero risk: they sit on top of tools you likely already run, and removing them later costs nothing. Orca and OpenMontage ask more, a full application install or a paid video model subscription, but nothing that touches the rest of your setup. Anydoc and Archify are narrow, purpose built tools that either solve a real problem you have today or sit unused, with little in between. DeepSeek Harness and Omarchy sit apart from the other eight: one asks you to learn a new plugin architecture instead of prompting a familiar one, the other asks you to change your operating system, and Chase's own framing, "a level above anything else we'll talk about here," is the right amount of caution to bring to that pair. Claude of Tanks needs no decision at all; it is worth a look for exactly what it is, an existence proof, not a workflow tool.