GitHub Trending this week: the AI repos, checked

GitHub Trending is GitHub's own list of the repos that get a lot of attention right now. "See what the GitHub community is most excited about this week," the page says. A repo is a project on GitHub: the code, the files and their history. Every week we take the AI repos from that list and check the facts. What does each one do, and who made it? How fast did it grow? Under which licence may you use it? We read the READMEs, but we do not run the code. The ones that stand out get a full test in their own article.

The AI repos on this week's list

These are the AI repos on GitHub's weekly list, in the order GitHub showed them. A skill or plugin below is an add-on with instructions or tools for an AI coding helper such as Claude Code, Codex or Cursor.

[mattpocock/skills](https://github.com/mattpocock/skills) (#2): +9,179 stars this week, 285,119 in total. Licence: MIT. Matt Pocock's set of skills for AI coding agents, "straight from my .agents directory". We tested two of them in Claude Code. The skill diagnosing-bugs helped the small model Haiku 4.5 fix more of a bug: 155 → 160 of 186 checks. The skill tdd left Sonnet 5.5 at a perfect score, but it took twice as long. Read our test.

[thedotmack/claude-mem](https://github.com/thedotmack/claude-mem) (#4): +4,035 stars this week, 99,435 in total. Licence: Apache-2.0. A memory for Claude Code and other agents. It records what the agent does in a session, sums it up with AI and gives the useful parts back next time, the README says. The installer asks you to sign in to claude-mem's own online service, with a 30-day free trial. The README also explains how to skip that sign-in.

[heygen-com/hyperframes](https://github.com/heygen-com/hyperframes) (#5): +4,081 stars this week, 60,526 in total. Licence: Apache-2.0. From HeyGen, a company that makes AI video. HyperFrames turns a web page (HTML, CSS and animations) into an MP4 video. It comes with 21 skills, so you can ask an AI agent for something like "a 10-second product intro" and let it build the video.

[mvschwarz/openrig](https://github.com/mvschwarz/openrig) (#6): +2,068 stars this week, 6,806 in total. Licence: Apache-2.0. Builds a team of AI coding agents, for example Claude Code and Codex together. You describe the team in a YAML file and start it with one command. The README lists macOS and Linux, plus a tool called tmux. One ready-made setup lets the agents skip permission prompts. That is handy, but it has real risks: we tested what skipping permissions does.

[cursor/plugins](https://github.com/cursor/plugins) (#7): +1,110 stars this week, 10,807 in total. Licence: none listed. Cursor's official plugins, like Gmail, GitHub, Playwright and Salesforce, plus Cursor's own workflow plugins. There is no licence file. Using the plugins inside Cursor is what they are made for. But without a licence, nobody may copy, share or change the code without permission, GitHub's own docs explain.

[earthtojake/text-to-cad](https://github.com/earthtojake/text-to-cad) (#8): +2,369 stars this week, 19,079 in total. Licence: MIT. Lets an AI agent design 3D parts. It makes STEP, STL, GLB and 3MF files, checks if a part can be made, draws technical drawings and links to 3D printing and CNC services. It sends usage stats and crash reports by default, the README says, but not your files or prompts. You can turn that off.

[pingdotgg/t3code](https://github.com/pingdotgg/t3code) (#10): +2,367 stars this week, 26,874 in total. Licence: MIT. One app to control the coding agents on your computer: Claude Code, Codex, Cursor, Grok Build, OpenCode and Google Antigravity. There is also a phone app. It works with the subscriptions you already have. "Wait, what are you selling me?" the README asks. "Nothing."

[ranxianglei/billion-context](https://github.com/ranxianglei/billion-context) (#11): +259 stars this week, 822 in total. Licence: MIT plus an extra attribution term. Squeezes long agent chats into short summaries, so one session can keep going for days. It claims 5× fewer tokens (tokens are the small pieces of text an AI reads and bills for). The licence is MIT with one extra rule: a product that uses it must credit it. That is why GitHub cannot name the licence.

[tester-army/e2e](https://github.com/tester-army/e2e) (#13): +6,564 stars this week, 8,820 in total. Licence: Apache-2.0. A testing tool for websites and phone apps. You write a goal in plain words, and an AI agent clicks through your app to reach it. The next run repeats those steps without asking the AI again, until the app changes. This is the newest hit of the week: 74% of its stars came in these 7 days.

Not AI, but also on the weekly list: AnyPS5 at #1, a tool that ports PS5 programs to Linux and Windows. The others were Epic Games' raddebugger, Flutter and openGym, a gym tracker you host yourself.

New hit or old favourite?

GitHub shows how many stars a repo got this week. We divided that by its total stars. A high share means the repo is new or was just found by many people. A low share means it was already big and keeps growing.

• New or just found: tester-army/e2e (74%), billion-context (32%) and openrig (30%).
• Already big: mattpocock/skills (3% of 285,119 stars) and claude-mem (4%).

Stars say that people noticed a repo. They do not say that it works well. Some repos are also very busy: billion-context had 979 commits (saved changes) in 7 days, and t3code had 625.

Also on today's list

GitHub also has a daily list. On 11 October 2026 it showed these AI repos as well.

[morluto/rea](https://github.com/morluto/rea) (#1): +25,793 stars today, 82,755 in total. Licence: MIT. "Reverse Engineer Anything": one tool that lets your AI agent look inside apps and programs without their source code, and explain how a feature works. It got more stars in one day than any repo on the weekly list got in a whole week. Many app licences forbid reverse engineering, so check the rules before you point it at someone else's app.

[storytold/artcraft](https://github.com/storytold/artcraft) (#3): +3,222 stars today, 15,485 in total. Licence: Apache-2.0. "The IDE for artists": a desktop app for making AI images and videos, where you set up the scene in 2D or 3D first. The repo is from 2022, so this is an older project that is getting noticed again.

[cathrynlavery/diagram-design](https://github.com/cathrynlavery/diagram-design) (#4): +1,190 stars today, 49,391 in total. Licence: MIT. A skill that lets your AI agent draw diagrams as one HTML file, in 44 types, in your own colours and fonts.

[mksglu/context-mode](https://github.com/mksglu/context-mode) (#5): +178 stars today, 26,506 in total. Licence: Elastic License 2.0 (source-available). Keeps big tool results out of the AI's working memory (its context window), so long sessions last longer. The README claims a 98% cut. You may use, change and share it, but you may not sell it as a hosted service. That makes it source-available, not open source.

[hugohe3/ppt-master](https://github.com/hugohe3/ppt-master) (#9): +461 stars today, 59,906 in total. Licence: MIT. Makes real, editable PowerPoint files from a document or a topic. The README opens with paid sponsor messages, several with referral links to AI services.

[anthropics/knowledge-work-plugins](https://github.com/anthropics/knowledge-work-plugins) (#12): +625 stars today, 28,964 in total. Licence: Apache-2.0. Anthropic's own plugins for jobs like sales, legal, finance and marketing: 11 in its list. They are made for Claude Cowork and also work in Claude Code.

[multica-ai/andrej-karpathy-skills](https://github.com/multica-ai/andrej-karpathy-skills) (#13): +278 stars today, 218,566 in total. Licence: none listed. One CLAUDE.md file with four rules, based on a post by Andrej Karpathy, but not made by him. In our test it did not help: Haiku 4.5 went from 155 to 152 of 186 checks. Read our test.

Check the licence first

A licence says what you may do with code. 12 of the 16 AI repos here use MIT or Apache 2.0. Both are open source licences that let you use, change and share the code, also in paid products, as long as you keep the licence text.

Four are different:

• billion-context: MIT plus one extra rule. Credit the project if people can see your product.
• context-mode: Elastic License 2.0. You may use and change it, but not offer it as a hosted service.
• cursor/plugins and andrej-karpathy-skills: no licence listed. Then normal copyright applies, and nobody may copy, share or change the code without permission, GitHub says.

Reading a repo is always fine. Check the licence before you copy code into your own project.

Repos from this list we tested ourselves

Two repos on this week's lists already had a full test from us. Both ran in Claude Code with a clean setup, on the same three secret coding jobs, three tries each.

• [Matt Pocock's skills](/news/matt-pocock-skills-claude-code-test): Haiku 4.5 155 → 160 of 186 checks with the skill. Sonnet 5.5 stayed at 186, but took 67 seconds per job instead of 33.
• [Karpathy's CLAUDE.md](/news/karpathy-claude-md-skills-test): Haiku 4.5 155 → 152, Sonnet 5.5 186 → 186. No gain.

So far, a lot of stars has not meant a big gain in our tests. We pick new repos from this list for our next tests.

How we checked

We saved GitHub's Trending page, the weekly and the daily list, on 11 October 2026 at about 09:26 UTC. The lists change during the day. We picked the AI repos and AI coding tools. Then we looked up each repo with the GitHub API (GitHub's data service for programs) at about 09:28 UTC: stars, licence, creation date and commits on the main branch in the last 7 days. That is why our star counts can differ a little from the screenshot. Where GitHub could not name a licence, we read the LICENSE file and the README.

We read each README and wrote down what the makers say. Their claims stay claims: we did not install or run any of this code for this page. GitHub no longer shows who starred a repo, so we could not check for fake stars.

The charts and the diagram were made by us. The screenshot is real and cropped. None of the images was made by AI. Next update: the week up to 18 October 2026.