12 free GitHub repos for Claude Code worth installing
Twelve free GitHub repos for Claude Code that gained 50,000 stars in a single week, plus three bonus picks. What each one does, the honest caveats, and which to install first.
10 min read
Every Sunday a new batch of GitHub repos trends for Claude Code, and most of them are not worth your time. This week twelve of them gained 50,000 stars between them, and a short list of three more rode along as bonus picks. None of them were tested by the person who compiled the list, so everything below is taken from each repo's own README, demo clips and documented output. Treat it as a scouting report, not a review.
The repos fall into a few groups: safety (scanning the skills you already installed), control (driving Claude Code from your phone or your desktop apps), knowledge (modern web guidance and a design system Claude can look up instead of guessing), media (video editing on your machine), quality (fresh agents that review your spec), and one that turns a single sentence into a playable browser game.
The short version
If you install only one thing, make it the scanner. It runs free on the skills you already have, and the numbers behind it are the reason the whole list exists: out of 31,000 skills NVIDIA looked at, 26% had a vulnerability and 5.2% showed likely malicious intent. Everything else on this list is a nice-to-have next to that.
1. NVIDIA SkillSpector: are your skills safe?
Every skill you add to Claude Code runs with your permissions, your files and your keys. SkillSpector is NVIDIA's free scanner for those skills: 71 checks, and the basic scan needs no API key. You install it with one line using uv, then point it at a folder or a GitHub link before you install anything.
- Scores a skill 0 to 100 with a severity and a recommendation (do not install, review, and so on).
- Scans a local folder or a GitHub link, so you can check a skill before installing it.
- Basic scan is free and pattern-based; no API key required.
- Honest limits: pattern matching can flag things that are fine, and a clean score is not a guarantee. It is a smoke alarm, not a fire department.
2. T3 Code: Claude Code from your phone
Claude Code is twenty minutes into a task and you go make coffee. You come back and it has been waiting for a yes since the minute you left. T3 Code fixes that. It is one app that drives the agents already on your computer (Claude Code, Codex, Cursor and a few more) and includes a phone app so you can approve the next step from the sofa.
- Drives the agents you are already signed into; sells no tokens.
- Each thread can work in its own copy of your project, so two tasks do not step on each other.
- Phone connects over your home network, Tailscale, or T3's own connect.
- MIT licensed. Try it once without installing using one command.
- Skip it if you only ask Claude quick questions.
3. Modern Web Guidance: stop 2019-style code
Claude learned to code from the internet, and most of the internet's code is old. Ask for a pop-up and you often get the 2019 answer: a pile of JavaScript for something the browser now does by itself. Modern Web Guidance is a set of skills from the Chrome team, with Edge, Firefox and MDN, that hands your agent the current way to do it.
- Hacker News thread reached 447 points this week.
- Try it without installing: search for what you are building and it finds the guide.
- Legacy scan points at a project and looks for old patterns to modernise.
- Guides cover layout, forms, performance, accessibility and the AI now built into the browser.
4. Google's Elephant-Goldfish Model
The agent that helped you design a feature is the worst one to review it, because it remembers every excuse you made. This plugin splits the job: the elephant holds all the context and writes the spec, then fresh goldfish agents with no memory of your chat read only the document and its sources and ask the rude questions.
- Design with an elephant, review with a goldfish.
- Drift report compares your documents with your code and reports where they disagree, so you can try it without writing a spec.
- Goldfish need fresh sub-agents; if your agent cannot create them, the review stops.
- A review does not replace tests or human judgement, and it uses more of your plan. Too much for a one-line fix, cheap for a feature you will live with for a year.
5. FFmpeg Skill: Claude edits your video
FFmpeg is the free tool behind half the video apps you use, and it is famous for commands nobody can remember. This skill gives Claude Code 42 of them ready to use: cut, join, captions, remove silence, turn wide video into vertical, all on your machine, with no cloud and no API keys.
- One-sentence requests work: take the interview, keep these two parts, make it exactly 60 seconds for Reels.
- The agent checks the file, cuts, reframes, exports, then checks its own work and reports back (59.98 seconds, 1080 by 1920, 30 fps).
- A doctor command checks the setup.
- Caveat: it can report not OK because of a single missing filter while most tools still work, and it will not pick the good parts for you. You say where to cut; it does the cutting.
6. Plannotator: mark up Claude's plan in your browser
When Claude shows you a plan, do you read it or do you press yes? The terminal is a bad place to read. Plannotator moves the plan to your browser, where you can comment and cross things out, then send your notes back to the agent. It reviews code changes the same way.
- Inbox lets several agents leave you questions that you answer when you are free.
- It was on GitHub's trending list both today and this week.
- Install caution: the default script sets up skills, hooks and commands for your agents. Start with the minimal flag, which installs only the app.
7. HA-MCP: Claude runs your smart home
This one is not a coding tool. It connects Claude Code to Home Assistant, the free software people use to run smart lights, plugs and sensors. Type create an automation that turns on the porch light at sunset and Claude builds it.
- 2,300 stars this week.
- Installs from inside Home Assistant; you copy one link into Claude.
- Safety: choose an account just for this, with local access only. You are giving an AI the keys to your front door light, maybe not the lock.
8. Billion Context: one session for weeks
This is the one the author said they would not put on their main account, and also one of the most interesting. Billion Context promises a single Claude Code session that runs for a month, using five times fewer tokens. It sits between your agent and the model and compresses the conversation on the way through.
- The author published a study: 4.5 months, 174,000 model calls, nearly 19 billion tokens. Those are the author's numbers and were not checked.
- A tool in that position sees everything: your code, your prompts, and it rewrites what the model receives.
- MIT licence plus one extra term about giving credit, which is why GitHub cannot label it.
- Try it first on a side project with a key that has a spending limit.
9. Claude Video: /watch any video
Claude cannot watch a video. This skill gives it one command that gets close: /watch followed by a link. It pulls the captions, grabs frames with timestamps, and hands them to Claude, so you can ask what is on screen at a certain moment in someone else's video or your own screen recording.
- Runs in Claude Code only. It does not work in Claude chat or Cowork.
- The last update was in September; it is back on the trending list this week.
- Install shows in three screenshots in the Claude desktop app: customise plugins at the marketplace.
10. Meta Astryx: the week's biggest gainer
13,600 stars in seven days, from Meta. Astryx is the design system they have used inside the company for eight years, across 13,000 apps: a box of finished parts (buttons, forms, tables, menus), over 150 of them, all matching and free.
- Ships with a command-line tool and docs written for agents, so Claude looks up the real part instead of guessing one.
- Seven themes included.
- Still in beta and needs React 19. Start here for a new project; wait for an old one.
11. Agent Desktop: Claude clicks your apps
Agent Desktop lets an agent use the apps on your computer, not by looking at screenshots but by reading the same hidden map of buttons that screen readers use. Its quick start asks what is in Finder and gets back 59 elements, every button and folder tagged, which the agent then clicks by tag.
- Works on apps that have no other way in for an AI.
- On a Mac your terminal needs the accessibility permission, which is a big permission.
- Snapshots can contain private things from your screen; keep them out of public logs.
12. Claude Code Game Studios: one sentence to a playable game
The last one is for people who do not code. Open Claude Code in a copy of this repo and type one sentence. Their example is a fishing game where the rare fish only bite at dusk. It asks a few questions (how should it feel, how hard, what is the goal), writes a one-page brief, waits for your go-ahead, then builds a browser game that opens with a double click.
- Tests the game in a hidden browser before handing it over.
- The handover note ends with the line: I love not yet played by a person.
- Their README has no screenshot of a finished game, so none is shown here and none is faked.
- Costs real usage; ask it to build at medium for a cheaper, thinner result.
Three bonus picks
- Caveman (120,000 stars): makes your agent talk like a caveman to save tokens. Their example goes from 63 tokens to 20 with the same fix. A joke that turned out to work.
- Claude Trading Skills: a 15-minute daily market check, risk rules and a trade journal as Claude skills. The author states clearly it is not a signal service and not financial advice.
- LoopForge (about 1,400 stars, from Tencent): runs coding agents in steps and stops for your approval between them.
Which to install first
- Install SkillSpector and run it on the skills you already have. The basic scan is free.
- If you run long tasks, add T3 Code so you can approve from your phone.
- If Claude builds your websites, install Modern Web Guidance before your next project.
- For video work, try the FFmpeg Skill on one real clip.
- Leave Billion Context on a side project with a spending-limited key until you have read it yourself.
Run it on the skills you already have. The basic scan is free.
If you found this useful, browse the rest of the YesIntelligent articles for more practical guides on creator tools, AI agents and reading engagement honestly.
Related articles
Every free tool YesIntelligent offers, and when to use each
Nine small tools that cover the repetitive jobs of a creator's week: pulling the words out of a video, checking a post before it gets cut off, reading an engagement number honestly, turning a good line into an image, finding prospects, and rewriting stiff AI copy.
YouTube transcript generator: use cases and why ours
A YouTube transcript is the fastest way to turn a video you admire into text you can search, quote, and reuse. Here are the jobs it does best, and why ours stays out of your way.
Instagram Reel transcript generator: use cases and why ours
Reels are short, fast, and often uncaptioned. A transcript turns each one into text you can study in a minute, even when nothing is written on screen.