AI coding repos we track
The repos we track for coding agents, agent skills, memory, MCP servers, local models and self-hosting. Stars, activity and releases refresh from GitHub every morning. Most are open source. Where one is not, our note says so.
Synced 7 hours ago
50 repos
| # | Repository | Category | Stars | Last push | Release | Licence |
|---|---|---|---|---|---|---|
| 1 |
Skills that make a coding agent write a spec and a plan first, then build with tests and reviews. Install steps for 16 agents. One main author. Optional visual companion loads a logo from their site. |
Claude Code and agent skills | 321,451 | 2 weeks ago | v7.0.2 | MIT |
| 2 |
Open-source coding agent for the terminal, with a desktop app in beta. Ships a full-access build agent and a read-only plan agent. Large team, frequent releases. |
Coding agents | 227,878 | 7 hours ago | v1.19.6 | MIT |
| 3 |
Anthropic's example skills and the Agent Skills spec, with the docx, pdf, pptx and xlsx skills. The README calls them demonstrations and says Apache 2.0, but the repo has no licence file or releases. |
Claude Code and agent skills | 197,323 | 2 days ago | not stated | |
| 4 |
Runs open models on your machine behind a local REST API, and can launch Claude Code, Codex and other agents against them. Cloud models are offered too. Python and JavaScript libraries. |
Local models | 187,567 | 14 hours ago | v0.40.16 | MIT |
| 5 |
Not the source code. This repo holds the changelog, issue tracker, examples and plugin docs for Claude Code. The licence is all rights reserved, under Anthropic's commercial terms. |
Coding agents | 156,201 | 8 hours ago | v2.1.328 | not stated |
| 6 |
A CLI (specify) that adds spec-driven steps to a coding agent: constitution, spec, plan, tasks, implement. Needs Python 3.11+ and uv. Its README warns that community extensions are not audited. |
Context, cost and workflow | 150,535 | 16 hours ago | v1.4.2 | MIT |
| 7 |
OpenAI's terminal coding agent, written in Rust, Apache-2.0. Sign in with a ChatGPT plan or an API key. Alpha builds ship almost daily next to the stable releases. |
Coding agents | 135,815 | 7 hours ago | rust-v0.170.0 | Apache-2.0 |
| 8 |
C/C++ engine that runs GGUF models locally, with a CLI and an OpenAI-compatible server. Pulls models from Hugging Face. Builds are tagged several times a day. |
Local models | 135,633 | 8 hours ago | v1.0.0 | MIT |
| 9 |
Desktop app that runs Claude Code, Codex and other agents in parallel, each in its own git worktree, with diff review and a mobile app. A company product with cloud options. |
Context, cost and workflow | 129,014 | 7 hours ago | v1.4.263 | MIT |
| 10 |
A minimal agent harness you extend yourself. Leaves out sub-agents and plan mode on purpose. No built-in permission system: it runs with your user's rights. Needs Node.js 22.19 or newer. |
Coding agents | 122,802 | 13 hours ago | v1.5.2 | MIT |
| 11 |
A skill and a local proxy that shorten what a coding agent says and reads. Six parts, each with a switch. The CLI sends usage stats by default. One maintainer. Savings figures are its own benchmarks. |
Claude Code and agent skills | 121,361 | 1 day ago | v4.1.0 | Apache-2.0 |
| 12 |
25 skills and 9 slash commands that walk a coding agent through spec, plan, build, test, review and ship. Installs as a Claude Code plugin or with the skills CLI. One main author. |
Claude Code and agent skills | 119,004 | 16 hours ago | 0.6.17 | MIT |
| 13 |
A second agent takes notes and loads them into your next session. Memory stays local. The hosted observer is a 30-day trial that asks for a card. Sign-in is optional. Analytics on by default. |
Agent memory | 118,283 | 8 hours ago | v16.0.0 | Apache-2.0 |
| 14 |
Design commands for a coding agent (critique, audit, polish) and a detector with 62 fixed rules that runs with no LLM or API key, in the agent, in CI or as a Chrome extension. One main author. |
Claude Code and agent skills | 113,014 | 13 hours ago | skill-v4.10.0 | Apache-2.0 |
| 15 |
No longer for individuals. Its README says free and Google AI Pro or Ultra access moved to Antigravity CLI on June 18, 2026. It now serves Gemini Code Assist licences and is not tuned for Gemini 4. |
Coding agents | 108,933 | 10 hours ago | v0.69.0 | Apache-2.0 |
| 16 |
Python library and server for serving LLMs at volume, with an OpenAI-compatible API and an Anthropic Messages API. Built for GPUs, with CPU and other hardware support. Large community project. |
Local models | 97,192 | 7 hours ago | v0.34.0 | Apache-2.0 |
| 17 |
Self-hosted control centre that runs its own agent, or Claude Code, Codex and others, locally, in Docker or on a server. Without the Docker sandbox the agent has full access to your files. |
Coding agents | 93,791 | 11 hours ago | v1.42.0 | MIT |
| 18 |
Seven reference MCP servers (fetch, filesystem, git, memory and more), meant as teaching examples for building your own. Older servers such as GitHub and Postgres moved to an archive repo. |
MCP servers | 92,093 | 13 hours ago | v2.5.0 | not stated |
| 19 |
CLI that compresses command output (tests, diffs, logs) before a coding agent reads it. The CLI is Apache-2.0, but its engine is BUSL-1.1, source-available. Telemetry is off by default. |
Context, cost and workflow | 87,123 | 11 hours ago | v1.2.0 | Apache-2.0 |
| 20 |
Compresses tool output, logs and history before they reach the model, as a Python library, a proxy or an agent wrapper. Runs locally. An anonymous beacon is on by default. Savings are its own figures. |
Context, cost and workflow | 80,117 | 12 hours ago | v0.44.0 | Apache-2.0 |
| 21 |
AI pentesting agents that attack a running app and write proof-of-concept exploits. Needs Docker and your own LLM key. Its README limits use to systems you own or may test. A hosted version exists. |
AI security | 76,573 | 7 hours ago | v1.7.1 | Apache-2.0 |
| 22 |
A skill that runs a six-phase security audit of a codebase with isolated sub-agents, and makes a fresh agent try to disprove each finding. Prompt files only, no releases. From Cloudflare. |
Claude Code and agent skills | 75,448 | 2 months ago | MIT | |
| 23 |
Open-source coding agent as a CLI, a desktop app, a VS Code extension, a JetBrains plugin and an SDK. Has Plan and Act modes and works with many model providers. Apache-2.0, from Cline Bot Inc. |
Coding agents | 71,760 | 12 hours ago | v4.1.26 | Apache-2.0 |
| 24 |
A skill that rewrites AI-sounding text against 26 patterns from Wikipedia's guide to AI writing. Its README says it does not try to pass AI detectors, and that they still flag most of its output. |
Claude Code and agent skills | 70,894 | 1 week ago | v4.0.0 | MIT |
| 25 |
A skill that researches a topic across Reddit, Hacker News, Polymarket and GitHub with no setup, then X, YouTube and more once you add keys. All keys are optional. Several releases on some days. |
Claude Code and agent skills | 70,215 | 1 day ago | v4.3.2 | MIT |
| 26 |
Memory layer for agents, as a library, a self-hosted server or a hosted cloud. Needs an LLM, OpenAI by default. Its README says the benchmark scores are for the cloud, not the open-source SDK. |
Agent memory | 69,220 | 1 day ago | ts-v3.3.1 | Apache-2.0 |
| 27 |
A skill that draws 50 diagram types as one self-contained HTML and SVG file, and can copy your site's colours and fonts. No server, accounts or telemetry. Diagrams load their fonts from Google Fonts. |
Claude Code and agent skills | 68,281 | 12 hours ago | MIT | |
| 28 |
Self-hosted alternative to Heroku, Netlify and Vercel. Deploys apps, databases and 360+ one-click services to any server you reach by SSH. Free with no paywalled features. A paid cloud version exists. |
Self-hosting | 66,235 | 17 hours ago | v4.7.5 | Apache-2.0 |
| 29 |
Feeds current, version-specific library docs to a coding agent, as an MCP server or a CLI with a skill. A free API key raises rate limits. Only the server is open: the backend and crawler are private. |
MCP servers | 65,095 | 10 hours ago | @upstash/context7-... | MIT |
| 30 |
Gateway and Python SDK that call 100+ LLM providers in the OpenAI format, with virtual keys and spend tracking. MIT, except the enterprise folder, under its own licence. Dev builds ship almost daily. |
Evals and observability | 63,515 | 7 hours ago | v1.110.1 | not stated |
| 31 |
Agent memory server that aims to make agents learn over time, not only recall chats. Docker is the recommended install. Needs an LLM key, 25+ providers supported. MIT. Accuracy claims are its own. |
Agent memory | 61,892 | 14 hours ago | v1.2.1 | MIT |
| 32 |
Google's MCP server and CLI that let a coding agent drive and inspect a live Chrome: traces, network, console, screenshots. Usage statistics go to Google by default, with a flag to opt out. |
MCP servers | 58,249 | 7 hours ago | chrome-devtools-mc... | Apache-2.0 |
| 33 |
AI pentester that reads your source code, then runs real exploits against the running app. Needs Docker. AGPL-3.0. Its README says to use it only on systems you may test, and never on production. |
AI security | 52,057 | 13 hours ago | v3.5.1 | AGPL-3.0 |
| 34 |
Desktop app and engine that run open models such as Qwen on your own computer, with installers for Windows, Linux and Mac. No account needed. MIT. Almost all commits come from one author. |
Local models | 45,721 | 14 hours ago | v0.2.16 | MIT |
| 35 |
Microsoft's MCP server for browser automation through Playwright's accessibility tree, with no vision model. Its own README says coding agents may do better with the Playwright CLI and skills. |
MCP servers | 39,869 | 18 hours ago | v0.0.86 | Apache-2.0 |
| 36 |
Self-hosted PaaS for apps and databases, built on Docker and Traefik, with 600+ templates and multi-node through Docker Swarm. Apache-2.0 except a proprietary folder. A cloud version exists. |
Self-hosting | 38,604 | 9 hours ago | v0.31.4 | not stated |
| 37 |
Tracing and evals for LLM apps and agents, self-hosted or as a cloud. MIT except the enterprise folders. Part of ClickHouse since January 2026. Several releases a week. |
Evals and observability | 37,612 | 7 hours ago | v4.82.0 | not stated |
| 38 |
GitHub's official MCP server: lets an agent read repos, manage issues and pull requests, and inspect Actions runs. Runs as a remote server with OAuth or locally with a personal access token. MIT. |
MCP servers | 34,578 | 14 hours ago | v2.2.0 | MIT |
| 39 |
Plugin that runs tool calls in a local sandbox and keeps the full output in SQLite, so only the needed part reaches the agent. Free, but Elastic License 2.0, not open source. Claims zero telemetry. |
Context, cost and workflow | 34,286 | 9 hours ago | v2.2.1 | not stated |
| 40 |
Python memory platform that turns documents, tickets and agent runs into a knowledge graph for agents. Needs Python 3.10 to 3.14. Apache-2.0. A managed cloud version exists. |
Agent memory | 33,396 | 13 hours ago | v1.6.10 | Apache-2.0 |
| 41 |
Python framework that builds knowledge graphs which track how facts change over time, for agent memory. Self-hosted only. From Zep, whose own platform is the managed version. Apache-2.0. |
Agent memory | 32,934 | 8 hours ago | v0.31.2 | Apache-2.0 |
| 42 |
Memory and context engine for AI apps and coding tools, as a cloud API or a self-hosted server with local embeddings. Telemetry can be switched off. MIT. Its benchmark claims are its own. |
Agent memory | 31,669 | 12 hours ago | server-v0.0.9 | MIT |
| 43 |
MCP toolkit that lets a coding agent find, edit and refactor code by symbol, not by line. Needs only uv. GPL-3.0 for the app, MIT for the language layer. A paid JetBrains plugin exists. |
MCP servers | 31,519 | 17 hours ago | v2.1.0 | not stated |
| 44 |
CLI that packs a whole repository into one file for an AI model, with a token count and a check for secrets such as API keys. MIT. Mostly one author. A release every month or two. |
Context, cost and workflow | 29,327 | 1 day ago | v1.19.0 | MIT |
| 45 |
CLI and library to test prompts, compare models side by side and red-team LLM apps. Runs locally, and most providers need an API key. Part of OpenAI now, still MIT by its own README. |
Evals and observability | 27,437 | 13 hours ago | code-scan-action-0... | MIT |
| 46 |
NVIDIA's scanner that checks an agent skill for vulnerabilities and malicious patterns before you install it. Works best with an OpenAI-compatible LLM endpoint. Installs third-party software. |
AI security | 26,846 | 12 hours ago | v2.13.6 | Apache-2.0 |
| 47 |
Python framework for testing LLM apps, built like Pytest: metrics for agents, RAG and chatbots. Apache-2.0. From Confident AI, whose hosted platform the README promotes for production use. |
Evals and observability | 19,551 | 8 hours ago | python-v4.2.4 | Apache-2.0 |
| 48 |
Reads the session files your coding agents already write and shows cost by tool, model, project and task. No account, API key or proxy. Terminal, web and desktop views. MIT. Mostly one author. |
Context, cost and workflow | 11,866 | 13 hours ago | desktop-v0.9.28 | MIT |
| 49 |
NVIDIA's scanner that probes an LLM for prompt injection, data leakage, hallucination and jailbreaks. Free, Apache-2.0. You need access to the model you test. A release every month or two. |
AI security | 10,094 | 16 hours ago | v0.17.1 | Apache-2.0 |
| 50 |
Laravel's official MCP server that gives a coding agent Laravel-specific context. Needs PHP 8.2 and Laravel 12 or 13. The README is short and points to the Laravel docs. MIT. |
MCP servers | 3,692 | 5 days ago | v2.10.4 | MIT |
| No repos match. | ||||||
A line marked in blue is our own note, written after we opened the repo. A Tested badge appears only where a post of ours ran it. Other rows show GitHub's own description.
Star growth over 30 days will appear once our own daily record is that long. It starts on November 16, 2026.
A repo with no push in 90 days is marked slow, after 365 days dormant, and archived when its owner archived it. "Not stated" means GitHub could not match the licence file to a known licence. Stars show attention, not quality.