Claude Code GitHub Repos: Manim, Remotion, Blender MCP and More
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I count down five free GitHub repos that turn Claude Code from something that writes code into something that makes videos, builds 3D scenes, and uses your whole computer. For each one I cover what it is, how it works step by step, who built it, and the catch. It's for anyone who already uses a coding agent and wants to know which of these are worth putting on their machine.
What I cover in this video
- 0:00 The five repos, and the one that went from about 4,000 stars to over 20,000 in nine days
- 0:46 Number 5: Manim, the engine behind the 3Blue1Brown look
- 2:17 Number 4: Remotion Skills, and the license most people skip
- 4:34 Number 3: Lightpanda, a headless browser built for AI agents
- 6:53 Number 2: MCP for Blender, Claude building scenes in your live Blender
- 9:11 Number 1: agent-desktop from Vercel Labs
- 12:15 Which ones I'd actually install, plus a bonus repo
The two video repos: Manim and Remotion Skills
Manim
Manim has over 42,000 stars and an MIT license. Grant Sanderson built the original for his 3Blue1Brown channel, and this is the community fork from 2020, which is the one you want. Here's how Claude uses it:
- Claude writes a Python file with a scene in it.
- Every shape (a square, a circle, text, a graph) is an object Manim calls a mobject.
- Claude tells the scene what to play: create the square, transform it into a circle, fade it out.
- You render it to an MP4 with the low quality flag.
- Claude saves the last frame with
-s, checks it, and fixes whatever is off.
manim -ql
There's no timeline and nothing to drag around. It's all text, which is exactly what a coding agent is good at. On a Mac you install Cairo and pkg-config with Homebrew, then install Manim with uv.
brew install cairo pkg-config
The catch: the install can fight you a little, the readme says it's mid refactor and not taking new feature contributions, and the look is math class. Great for diagrams, not for a flashy promo. We built our own explainer engine, but if I were starting today, I'd start here.
Remotion Skills
Remotion Skills has over 6,000 stars and comes straight from the Remotion team. Remotion makes videos out of React code, and these are its official agent skills: 13 folders and about 180 markdown files that teach your agent how Remotion wants to be used.
npx skills add remotion-dev/skills
You type something like /remotion create make a promo video for a record store. It sets up a blank project and opens an interactive preview first instead of rendering. Claude writes a React composition using useCurrentFrame, so every piece of the video asks what frame it's on, and interpolate turns that frame number into a position or opacity. At frame 0 the title is invisible, and by frame 30 it's fully there.
The big rule in the skill is that CSS transitions and animations won't render correctly. The renderer opens headless Chrome, takes a picture of every frame, and stitches them into an MP4, so anything running on its own clock comes out wrong. I built my intro and chapter card overlays in Remotion without these skills, and the render skill basically corrected our custom browser path setup.
The license: Remotion is free for individuals and teams of up to three. If your company has four or more people total, you need a company license at $25 a month per person writing Remotion code, and per their FAQ that includes writing it with an AI agent. Check the pricing page for minimums.
Agents on the web and in 3D: Lightpanda and MCP for Blender
Lightpanda
Lightpanda has over 41,000 stars. It's a headless browser built from scratch for AI agents, written in Zig, with no Chrome code inside. libcurl fetches the page, a parser called html5ever builds the DOM, and V8 runs the scripts. Then it skips what Chrome spends its memory on: no layout, no painting, no graphics card. It speaks CDP, so existing Puppeteer or Playwright scripts can point at it. There's also a built in MCP server and a lightpanda agent command that saves what it did as a script you can replay later without a model.
By their own benchmark on their own demo site, running 25 browsers at a time it was about nine times faster with about a sixteenth of the memory. One at a time, it's about one and a half times faster. So the real win is memory when you run lots of agents on one server. Since the end of August it can take a screenshot, but it's a picture of the page's text with no layout or styling. If your agent works by looking at the page, this isn't it.
The catch: it's AGPL, some sites won't work without layout, there's no native Windows version (it runs in WSL), and telemetry is on by default, so set the disable flag.
MCP for Blender
MCP for Blender has over 33,000 stars and was renamed from Blender MCP in September. It's a community plugin, not made by Blender, that lets Claude Code, Codex, Cursor, and others build and edit 3D scenes in your live Blender. An add-on opens a socket server on port 9876, the MCP server passes your agent's commands down it as JSON, and the main tool runs any Blender Python your agent writes. A tool called look shows the agent the viewport, camera, a sheet of angles, plus wireframe and X-ray views, so it loops: write the code, look, fix.
It can pull real models from Poly Haven, Sketchfab, and Poly Pizza and writes the license info onto the object. It's plugged into my Claude Code right now.
claude mcp add blender uvx mcp-for-blender
Then turn on the add-on inside Blender. The risk: by default it runs any Python the AI writes on your machine, and the socket has no authentication. Save your file first, turn on safe mode, and set disable telemetry to true.
agent-desktop: an AI that uses your whole computer
agent-desktop from Vercel Labs had about 4,000 stars on October 31st and is now over 20,000. It's computer use for any coding agent on Mac, Windows, and Linux, with no model inside. Your agent is the brain and this is the hands.
Instead of screenshots and guessing pixels, it reads the accessibility tree your OS keeps for screen readers. It takes a JSON snapshot where every button and field gets a tag like E8, your agent acts on the tag (there are 56 commands), and then it reads again, because a successful action doesn't prove the UI changed. If a tag goes stale, nothing is sent and it takes a fresh snapshot. It's headless by default, and each agent can get its own session, though actions take turns.
The warning label: on a Mac you give your terminal accessibility permission, and then any agent in that terminal can read and click anything in any app, including your email and messages. The creator's own advice is to run it inside a macOS virtual machine. I haven't run it and I won't on my main login.
Key takeaways
- MCP for Blender is the one I already use. Save first, safe mode on, telemetry off.
- Remotion Skills is a yes if you're solo or a team of three or fewer. Past that, budget for the seat.
- Manim is the cheapest way to get clean diagrams if you explain things.
- Lightpanda only makes sense for crawling or running lots of browser agents on a server.
- agent-desktop goes in a separate user or a VM, on boring apps first. Bonus: the one I run every day is librosa, which finds the beat so my shorts' scene changes land on it.
FAQ
Is Remotion free to use with Claude Code?
Remotion is free for individuals and teams of up to three people. If your company has four or more people, you need a company license at $25 a month per person writing Remotion code, and that includes writing it with an AI agent.
How do I connect Claude Code to Blender?
Install MCP for Blender with claude mcp add, then turn on the add-on inside Blender. Save your file first and turn on safe mode, since by default it runs any Python the AI writes.
Is agent-desktop safe to run on my main computer?
Once your terminal has accessibility permission, any agent in it can read and click anything in any app. The creator recommends running it in a macOS virtual machine, or at least use a separate user account and start with boring apps.
Can Lightpanda take screenshots for computer use agents?
It can take a screenshot, but it's the page's text flowed into blocks with no layout or styling. If your agent works by looking at the page, Lightpanda isn't the right fit.
Chapters
Transcript
Read the full transcript
0:00These five free GitHub repos turn Claude code from something that writes code into something that makes videos and uses your whole computer. And one of them went from about 4,000 stars to over 20,000 in 9 days. A free motion studio list on X passed 400,000 views and a near copy showed up in Japanese within a week. So in this video, we're counting down all five. The engine behind those three blue, one brown math videos. Official video skills from Remotion with a license most people skip. A browser built for robots that takes a screenshot of words. Claude building 3D scenes inside Blender. And an AI that uses your whole computer and why I'd give it a different one. Plus, which ones I'd
0:46actually install, and it's not all five. Let's jump right into it. So number five is Manom with over 42,000 stars and an MIT license. If you've watched a three blue, one brown video, that's the look. Grant Sanderson built the original for his own channel. So this is the community fork from 2020 and it's the one you want. So here's how it actually works. Step one, Claude writes a Python file with a scene in it. Step two, every shape is an object, which Manom calls a mob. A square, a circle, some text, a graph. Step three, Claude tells the scene what to play. Create the square, transform it into a circle, then fade it out. Transform moves the points of one shape until they become the other one.
1:32And then step four, you run manom-QL and it renders the whole thing into an MP4. Step five, Claude saves the last frame with - S, checks it, and fixes whatever is off. By the way, the flag for lowquality is QL. I appreciate the honesty. Now, here's what this is all about. There's no timeline and nothing to drag around. It's all text, which is exactly what a coding agent is good at. To install it on a Mac, it's brew install Cairo and pkg config, then uv manom. Now, the install can fight you a little, and the readme says it's in the middle of a major refactor, so they're not taking new feature contributions right now. And the look is math class. It's great for diagrams, not so much for
2:17a flashy promo. We don't use it. We built our own explainer engine, but if I were starting today, I'd start here. Real quick, 263 builders are already inside shipping school. In premium, I show you how I make every video on this channel with Claude Code. New course material drops every single day, plus live calls every week, over 500 hours of replays, and a full course library. Join free link below. So, number four is Remotion Skills. With over 6,000 stars, straight from the Remotion team, Remotion makes videos out of React Code, and it's at almost 72,000 stars. These are its official agent skills. 13 folders, about 180 markdown files that
3:02teach your agent how Remotion wants to be used. There's even one just for the license. You install them with npx skills addion-dev/skills. So, here's how it actually works. Step one, you type /ramotion create make a promo video for a record store. Step two, it sets up a blank project. Step three, it opens the preview first. The skill says don't render the video by default. They want to see an interactive preview. Step four, Claude writes react a composition and inside it a function called use current frame. Every piece of the video asks what frame is it right now? And then interpolate turns that frame number into a position or an opacity. At frame zero, the title is invisible and by frame 30, it's fully
3:49there. Step five is the big rule. The skill says CSS transitions and animations will not render correctly. That's the classic agent mistake because the renderer isn't playing your video live. It opens headless Chrome, takes a picture of every single frame, and stitches them into an MP4. So, anything running on its own clock comes out wrong. I built my intro and chapter card overlays in Remotion, and I haven't installed these skills yet. You can see our setup forces a custom browser path, and the render skill says, "Don't set that unless the user asks." So, I just got corrected by a markdown file. In fairness, the bundled Chrome wouldn't download for us. Here's something most people don't think about. Remotion is free for individuals and for teams of up to three people. If your company has
4:34four or more people in total, even if only one of them touches Remotion, you need a company license. $25 a month per person writing Remotion code. And that includes writing it with an AI agent. That's per their FAQ. So, check the pricing page for minimums. So, no, letting Claude write it doesn't get you out of the seat. The skills repo has no license file of its own. It comes out of the main Remotion repo and its license. So far, Claude is making videos. Now, it starts using computers, starting with the web. Number three is Light Panda with over 41,000 stars. It's a headless browser built from scratch for AI agents written in a language called Zigg, so there's no Chrome code inside it. So, here's how it actually works. Step one,
5:21lib curl fetches the page. Step two, a parser called HTML 5 ever reads it and builds the DOM, which is the tree of everything on the page. Step three, V8, the same JavaScript engine that's inside Chrome, runs the scripts. And then step four is the trick. It skips everything Chrome spends its memory on. So there's no layout, no painting, and no graphics card. Step five, it talks CDP, the same protocol Chrome uses, so your existing Puppeteer or Playright scripts just point at it. There's also a built-in MCP server for agents and a light panda agent command that saves what it did as a script you can replay later without a model at all. So, here are the numbers said honestly. By their own benchmark on their own demo site, 25 browsers at a
6:08time, it was about nine times faster with about a 16th of the memory. One at a time, it's about 1 and a half times faster, still with a fraction of the memory. So, the real win is memory when you run a lot of agents on one server. And here's the funny part. Since the end of August, it can take a screenshot. It's a picture of the page's text with no layout and no styling. Their own source code says a screenshot is the pages text content flowed into blocks. So if your agent works by looking at the page, which is how most computer use models work, this isn't it. If it reads pages and crawls at scale, it's great. So the company's angels include the founders of Mistl, Hugging Face, and Dust. And the pitch is a browser built assuming the user is a machine, not a
6:53human. Now, the catch is the license. It's AGPL. So, if you modify it and let people use your version over a network, you have to publish your changes. Some sites won't work because, like I said, there's no layout. There's no native Windows version, so on Windows, it runs inside WSL. And telemetry is on by default, so set the disable flag. So, number two is MCP for Blender with over 33,000 stars. It used to be called Blender MCP and it got renamed in September. It's a community plugin not made by Blender that lets Claude Code, Codeex, Cursor, and others build and edit 3D scenes in your live Blender. So, here's how it actually works. Step one, an add-on inside Blender opens a little
7:39socket server on port 9876. Step two, the MCP server talks to your agent and passes its commands down that socket as JSON. Step three, the main tool, execute Blender code, runs any Blender Python your agent writes. Step four, a tool called look. It shows the agent the viewport, the camera, a sheet of angles, even wireframe and X-ray views so it can check its own work. And then step five, it fixes what looks wrong and goes again. So it's a loop of write the code, look at it, fix it. So, it can also pull real models from Polyhaven, SketchFab, and Poly Pizza, and it writes the license info right onto the object. And a generate tool hooks into Tripo, Meshi, and others. With your own keys, you pay those
8:24providers directly, or they sell premium plans from $10 to $100 a month. It's plugged into my Claude code right now. You can see it right here. To be honest, my 3D clips still come from code with 3JS, but for building a real scene inside Blender, I think this is the one to install it in Claude Code. It's Claude MCP add Blender UVXMCP for Blender, then turn on the add-on inside Blender. One developer, Sidarth Ahuja, started it in March 2025. The Python package index shows over a million downloads last month, bots and mirrors included, and Verscell's CEO is listed as one of its supporters. Hold on to that because Verscell shows up in this video twice. Now, the risk. By default, it runs any Python the AI writes on your
9:11machine. The readme says in bold, always save your work before using it. So save your file first and turn on safe mode which checks every script and blocks things like reading files, running other programs or reaching the network. The socket also has no authentication. So anyone who can reach that port can run Python inside Blender and a small anonymous usage record is on by default. So set disable telemetry to true. So number one is agent- desk from Verscell Labs. In last week's AI roundup, I told you to try this. It had just under 8,000 stars. Now it's over 20,000. On October 31st, it had about 4,000. So that's almost five times in 9 days, which is insane. Their launch line is agent
9:58browser for your whole computer. It's computer use for any coding agent on Mac, Windows, and Linux. And there's no model inside. Your agent is the brain, and this is the hands. So here's how it actually works. Most agents take a screenshot, send the picture to the model, and guess which pixel to click. This one skips the picture. It reads the accessibility tree, which is the list of every button and field that your operating system already keeps for screen readers. Step one, it takes a snapshot. That's a JSON list of every button and field. And each one gets a short tag like E8. Step two, your agent acts on the tag like click E8. And there are 56 commands. Step three, it reads again because their docs say a successful action does not prove the UI
10:44changed. And then step four, if the app changed under it, the old tag goes stale, nothing gets sent, and it takes a fresh snapshot. It's also headless by default. It doesn't move your mouse or steal focus, so it mostly works in the background while you keep working, though some actions still need it on screen. So each agent can get its own session with its own snapshots and its own cursor. Chris Tate's launch post says, "Agents use your computer in parallel. The docs are more careful. The cursors travel at the same time, but their actions take turns." Their demo shows two agent cursors playing Garage Band together. It's a duet, and neither of them can play. So, who built it? Almost all of it is one developer, Lefir, with 230 commits, and he's at Versell now. Chris Tate, who made Agent
11:30Browser, launched it on October 30th, and that launch post got 1 and a half million views. And like I said, that's for cell twice. Now, this is the big one. On a Mac, you give the app that runs it, usually your terminal, accessibility permission. After that, any agent in that terminal can read and click anything in any app. So, think about what's in your other apps, like your email and your messages. Even the readme warns that snapshots and clipboard reads can contain private desktop content. Your whole computer is the selling point, and it's also the warning label. And the creator's own advice is to run Agent Desktop inside a Mac OS virtual machine. So do that or at least use a separate user account and start with boring apps. I haven't run it
12:15and I won't on my main login. It also can't see apps with no accessibility tree and version one is less than 2 weeks old. So which ones would I actually install? MCP for Blender. It's already in mine. Save your file first. Turn safe mode on and telemetry off. Remotion skills. Yes, if you're solo or a team of three or fewer. And past that, budget for the seat. Manom if you explain things because it's the cheapest way to get clean diagrams. Light Panda only if you're crawling or running lots of browser agents on a server. And agent desktop. Yes, in a separate user or a virtual machine on boring apps first. And a bonus because the one I actually run every day isn't on this list. It's Librosa. Since the end of October, every
13:01short I make runs Librosa to find the beat so the scene changes land right on it. So, which one is going on your machine first? And if it's Agent Desktop, tell me where you're running it because I want to know who's brave enough to hand it their main account. Comment down below. And if this one was useful, a like really does help. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Cloudflare Security Audit Skill, Graphify and 3 More Trending Repos
The five repos are Cloudflare's security-audit-skill (+15,072 stars), Graphify, ComfyUI, AI Memory and Ruflo (formerly Claude Flow), ranked by stars gained last week. I'd start with the security skill in quick mode on one folder, and add Graphify if your project is big.
Top Trending GitHub Repos: VoiceStudio, ai-job-search, Orca, Ghidra
The five repos right behind REA on GitHub Trending this week are VoiceStudio (free local voice cloning), ai-job-search (a Claude Code job hunt), Orca (a desktop app for running many coding agents), Ghidra (the NSA's reverse engineering tool), and HyperFrames (HTML to MP4 video).
5 GitHub Repos for Claude Code: REA, Matt Pocock Skills, HyperFrames
The five GitHub repos I've added to my AI building workflow are REA (Reverse Engineer Anything), Matt Pocock's skills, context-mode, Scroll World and HyperFrames by HeyGen. Most of them install as Claude Code plugins, or you can paste the repo link into your coding agent and ask it to set it up.
