The Repo Atlas.

Framework for building LLM applications and agents from models, tools and retrievers.

Graph runtime for agents that keep state, loop, and resume after an interruption.

Framework for connecting a model to documents: parsing, indexing, retrieval and agents over them.

Framework for agents that solve a task by exchanging messages with each other.

Fork of AutoGen run by a community organisation, continuing the earlier API.

SDK for calling models and plugins from C#, Python and Java applications.

Framework where each agent gets a role and a goal and the crew works through tasks.

Python framework for agents with memory, knowledge and tools, plus a runtime to serve them.

Formerly phidata.

Framework for building retrieval and agent pipelines out of swappable components.

Multi-agent framework built around role-playing conversations, used in agent research.

Multi-agent framework that hands out software company roles, from product manager to engineer.

Small library for agents that write and run Python as their actions.

Framework that lets agents built on different systems find each other and form teams.

No commits since October 2025.

Framework for running autonomous agents with a web console and a tool marketplace.

No commits since January 2025.

Experimental framework for an agent that writes and stores its own functions.

Its author describes it as a place to share ideas and says it is not meant for production.

Platform for building and running agents that work on a goal continuously.

Cloud sandboxes where an agent can run code and commands away from your machine.

Python library that lets an agent drive a real browser: click, type, read, extract.

Frontend toolkit for putting an agent inside the interface of a web or mobile app.

Runnable examples for comparing open agent frameworks on the same tasks.

Multi-agent framework that models a trading firm: analysts, researchers, a trader and a risk desk.

Engine that builds a simulated crowd of agents from seed material and reports what it predicts.

Self-hosted agent with persistent memory, skills it writes itself, and chat channels such as Telegram.

In this workshop: Runs the seven-agent fleet.

Personal assistant that runs on your own machine and answers through the chat apps you use.

Formerly Moltbot.

Self-hosted workspace for chat, agents, research, documents, email, notes and local models.

Job search agent that scans boards, scores each role against your CV and tailors the application.

iPhone app in which an agent builds web and iOS apps from a prompt and previews them.

Tool that plans and writes a social media campaign from a brand description.

No commits since February 2024.

Agent for machine-learning engineering tasks.

Hugging Face points users to HuggingChat. Archived by its owner. Last commit September 2026.

Coding agent for the terminal, with plugins, hooks and subagents.

In this workshop: Used daily, including for this site.

The agent is closed source. This repository holds plugins, examples and the issue tracker.

Coding agent for the terminal, written in Rust.

Terminal agent for Gemini models.

Terminal counterpart of Google's Antigravity editor.

In this workshop: Runs as one of the coding agents.

Closed source. The repository holds documentation, examples and a changelog.

GitHub Copilot's agent for the terminal.

Closed source under GitHub's own licence. The repository holds the installer and the issue tracker.

Terminal coding agent that works with many model providers.

Formerly sst/opencode.

Agent toolkit with one LLM API, an agent loop, a text interface and a coding agent.

Formerly badlogic/pi-mono.

Extensible agent that installs, runs, edits and tests code with any model.

Started at Block.

Pair programmer that edits files in your git repository and commits each change.

Agent that writes code, runs commands and browses inside a sandbox.

Formerly OpenDevin.

Coding agent aimed at open-weight models.

Coding agent for the terminal that reads and edits code, runs commands and fetches pages.

Replaces the archived Python kimi-cli.

Terminal coding agent for Qwen models.

Terminal coding agent from the makers of Bubble Tea.

Source-available. The licence turns into MIT two years after each release.

Coding agent available as an editor extension, a CLI and an SDK.

Coding agent for VS Code, JetBrains and the terminal.

Coding agent for editors and the terminal.

AI code editor.

Closed source. This repository is only an issue tracker.

Factory's coding agent for the terminal, the web and chat tools.

Closed source. The repository holds documentation.

Code assistant that answers from a search index of the codebase.

Sourcegraph stopped developing it in the open. This is the last public snapshot. Archived by its owner. Last commit August 2025.

Code search engine in Rust with natural-language questions about a repository.

Archived by its owner. Last commit December 2024.

Specification and documentation for the SKILL.md format that several agents share.

Anthropic's public repository of Agent Skills.

Works with Claude Code

Anthropic's directory of reviewed Claude Code plugins.

Works with Claude Code

OpenAI's catalogue of skills for Codex.

Works with Codex

Example Codex plugins, each with a manifest and optional skills, hooks and MCP servers.

Works with Codex

Skills for working with the Gemini API and SDKs.

Command-line tool that installs skills from any repository into the agents on your machine.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, OpenCode, Pi, Cursor, Copilot CLI, Kimi Code, OpenClaw

Skills framework and working method: brainstorm, plan, build with subagents, review.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, OpenCode, Pi, Cursor, Copilot CLI, Kimi Code

Large pack of skills, hooks, memory and rules for coding agents, installed as a plugin.

Offers a paid tier for private repositories.

Works with Claude Code, Codex, Antigravity, Hermes Agent, OpenCode, Cursor, Copilot CLI, Kimi Code, OpenClaw

One CLAUDE.md file of working rules, derived from Andrej Karpathy's notes on LLM coding mistakes.

Written by the community, not by Karpathy.

Works with Claude Code

Roster of specialist agent definitions, from frontend to community management, with an installer app.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, OpenCode, Cursor, Copilot CLI, Kimi Code, OpenClaw

Design reference skill for building interfaces across web and mobile stacks.

Works with Claude Code, Codex, Antigravity, Gemini CLI, OpenCode, Cursor, Copilot CLI

Skill that steers an agent away from generic frontend output.

Works with Claude Code, Codex, Cursor, Kimi Code

Design vocabulary and commands that make a coding agent's frontend work more deliberate.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, OpenCode, Pi, Cursor, Copilot CLI

Over 100 product management skills and commands, from discovery to launch.

Works with Claude Code, Codex, Gemini CLI, OpenCode

Skill that researches a topic across Reddit, X, YouTube, Hacker News and the web, then summarises.

Works with Claude Code, Codex, Gemini CLI, Hermes Agent, Cursor, Copilot CLI, OpenClaw

177 skills for scientific work, with connectors to over 100 research databases.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, Pi, Cursor

Writing skill that removes machine-sounding patterns from prose and repairs story structure.

Works with Claude Code, Codex, Antigravity

Registry of validated skills with an installer for many coding agents.

Works with Claude Code, Codex, Antigravity, Gemini CLI, OpenCode, Cursor, Copilot CLI

Plugin for spec-driven development: write the spec, plan the work, then implement.

Works with Antigravity, Claude Code

Workflow plugin for Claude Code made of agents, skills, hooks and commands.

Works with Claude Code, Codex, Cursor

Searchable directory of Claude skills, agents, commands, hooks and plugin marketplaces.

Works with Claude Code, OpenClaw

Collection of Claude Code agents, skills, commands, plugins, hooks and rules in one repository.

Works with Claude Code

Unofficial Python API and agent skill for Google's NotebookLM.

Codex skill that draws hand-sketched illustrations for Chinese-language articles.

Works with Codex

Runs and supervises teams of coding agents from planning to merge.

Formerly ComposioHQ/agent-orchestrator.

Works with Claude Code, Codex

Keeps coding agents running in tmux sessions and streams their terminals to any browser.

Runs on macOS and Linux.

Works with Claude Code, Codex, Antigravity, Gemini CLI, OpenCode, Pi

Lets agents in separate terminals message, watch and start one another.

Works with Claude Code, Codex, Antigravity, Gemini CLI, OpenCode, Pi, Cursor, Copilot CLI, Kimi Code

Calls Codex from inside Claude Code to review code or take over a task.

Works with Claude Code, Codex

Asks several models and coding CLIs the same question and shows the answers side by side.

Works with Claude Code, Codex, Antigravity, Cursor, Kimi Code

Lets one assistant hand subtasks to Codex, Claude Code, Cursor and Gemini.

No commits since October 2025.

Works with Claude Code, Codex, Gemini CLI, Cursor

Syncs skills, agents, MCP servers and plugins to every AI CLI on a machine from one source.

Works with Claude Code, Codex, OpenCode, Pi, OpenClaw

Manages skills, MCP servers, hooks, memory and rules for 15 agents from one place.

Its web mode runs on a headless Linux server.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, OpenCode, Cursor, Copilot CLI

Browses, searches and resumes the session history of coding agents.

macOS only.

Works with Claude Code, Codex, Antigravity, Hermes Agent, OpenCode, Pi, Cursor, Copilot CLI, Kimi Code, OpenClaw

Shows what each agent is doing and lets you approve actions from a Mac or an iPhone.

First released in September 2026. The Linux build is a beta and its Hermes plugin is Mac only.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, OpenCode, Cursor, Copilot CLI

Shows running coding agents as pixel-art characters at work in a small office.

Works with Claude Code

Keeps specs, tasks and memory in the repository so every coding agent follows the same project rules.

Works with Claude Code, Codex, Cursor

Blocks destructive commands and reads of secret files before an agent's tool call runs.

A guard, not a sandbox: it does not contain processes or watch the network.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, OpenCode, Pi, Cursor, Copilot CLI, Kimi Code, OpenClaw

Linter and language server for agent files: CLAUDE.md, AGENTS.md, SKILL.md, hooks and MCP configs.

Works with Claude Code, Codex, Gemini CLI, OpenCode, Cursor, Copilot CLI

Toolkit for adding programmable rules to what an LLM application may say and do.

Scanners for prompts and model output: injection, secrets, toxic content.

Archived by its owner. Last commit July 2026.

Finds bug and vulnerability patterns with rules that look like the code they match.

Read-only scan of a developer machine's packages and extensions for known supply-chain attacks.

Calls Zscaler AI Guard from NeMo Guardrails.

Demo that pairs NeMo Guardrails rules with Palo Alto Networks' runtime threat API.

No commits since March 2025.

Records what an agent does in a session, compresses it and feeds the relevant part into the next.

Works with Claude Code, Codex, Antigravity, OpenCode, Pi, OpenClaw

Memory layer that several agents can share for one person.

Works with Claude Code, Codex, Hermes Agent, Pi, Cursor, OpenClaw

Context database for agents that holds memory, knowledge and skills in one store.

Works with Claude Code, Codex, Hermes Agent, OpenCode, Cursor, OpenClaw

Compresses tool output, logs and files before they reach the model.

Works with Claude Code, Codex, OpenCode, Cursor, Copilot CLI, Kimi Code, OpenClaw

Skill and proxy that make an agent answer in terse language to spend fewer tokens.

Works with Claude Code, Codex, Gemini CLI, Copilot CLI

Hooks that trim what a coding agent reads: file hints, document conversion, caching.

Source-available. Commercial and team use need a paid licence.

Works with Claude Code, Codex, Gemini CLI, Copilot CLI

MCP server that indexes a codebase for semantic search by a coding agent.

Works with Claude Code

Builds a knowledge graph of a codebase and keeps it in sync for agents to query.

Works with Claude Code, Codex, Antigravity, Gemini CLI, Hermes Agent, OpenCode, Cursor, Copilot CLI

Specification and documentation of the protocol itself.

Reference server implementations maintained by the MCP project.

Community list of MCP servers, sorted by what they connect to.

Lets an assistant build and edit n8n workflows.

Community server for controlling Blender from an agent.

Not affiliated with the Blender Foundation.

Generates images and video through Higgsfield's hosted models.

A personal copy of the reference MCP servers.

No commits since February 2025.

C and C++ inference engine for LLMs on CPUs and GPUs, and the home of the GGUF format.

Downloads and runs open models locally behind one command and an HTTP API.

In this workshop: Serves the local models on the home GPU.

Inference server built for throughput, with batching and an OpenAI-compatible API.

Local server with an OpenAI-compatible API for text, vision, voice and image models. Runs without a GPU.

Packs a model and its runtime into a single executable file.

Formerly Mozilla-Ocho/llamafile.

Compiles LLMs into optimised TensorRT engines for NVIDIA GPUs.

Inference server for text generation models.

Archived by its owner. Last commit March 2026.

Self-hosted chat interface for Ollama and OpenAI-compatible servers.

A custom licence based on BSD-3-Clause.

Desktop and server app for chatting with your documents through local or hosted models.

Desktop chat app that runs models fully offline.

Self-hosted chat and agent workspace.

Formerly Lobe Chat. A custom licence based on Apache-2.0.

Desktop app for running local models on consumer hardware.

No commits since May 2025.

API layer for private document question answering on local models.

Formerly imartinez/privateGPT.

Open implementation of the NotebookLM idea: research and chat over your own sources.

Private chat with documents, images and video on local models.

Archived by its owner. Last commit October 2025.

Release repository for the DeepSeek-V3 language model.

No commits since August 2025.

Second document OCR model from DeepSeek.

Family of open multimodal models from OpenGVLab.

No commits since September 2025.

Vision-language assistant that answers questions about images.

No commits since August 2024.

Alibaba's first vision-language model family.

No commits since August 2024.

Fine-tunes and runs LLMs with low memory use, with a local interface.

One interface for fine-tuning over 100 language and vision models.

Fine-tuning driven by a single YAML config file.

Library of parameter-efficient fine-tuning methods such as LoRA.

Trainers for post-training language models with supervised, preference and reinforcement methods.

Optimisation library for distributed training and inference of large models.

Formerly microsoft/DeepSpeed.

Toolkit for parallel and distributed training of large models.

Removes refusal behaviour from a language model automatically.

8-bit and 4-bit quantization for PyTorch models.

Inference framework for 1-bit LLMs.

Half-quadratic quantization, which needs no calibration data.

Formerly mobiusml/hqq.

Quantization package based on the GPTQ algorithm.

Archived by its owner. Last commit April 2025.

4-bit quantization with the AWQ algorithm.

Archived by its owner. Last commit May 2025.

Framework for running language models through standard benchmarks.

Tests prompts, agents and RAG pipelines, and red-teams them for vulnerabilities.

Framework for evaluating models across many scenarios and metrics.

Platform for training, serving and comparing chat models. Chatbot Arena was built on it.

Collaborative benchmark of more than 200 tasks for language models.

Archived by its owner. Last commit July 2024.

Dataset of Chatbot Arena conversations with the human preference votes.

Framework for writing LLM pipelines as modules and letting an optimiser tune the prompts.

Language for constraining model output with grammars, regular expressions and control flow.

Tools for testing prompts across models and vector databases.

Guides, papers and notebooks on prompting, context engineering, RAG and agents.

Embedding database for search and retrieval in AI applications.

Vector database and search engine written in Rust.

Distributed vector database for large-scale similarity search.

Vector database that stores objects with their vectors and combines vector and keyword search.

Vector similarity search as a PostgreSQL extension.

RAG engine that parses complex documents before retrieval, with agent workflows.

Builds a tree index of a long document so a model can reason towards the right section.

Crawls websites and returns clean Markdown or structured data for LLMs.

Python scraping framework that adapts when a page's structure changes.

Library for building crawlers with plain HTTP or a headless browser.

Converts Office files, PDFs and other documents to Markdown.

Local document parser from the LlamaIndex team.

Command-line tool that lets an agent read and search X, Reddit, YouTube, GitHub and Bilibili.

Tracing, evaluation and prompt management for LLM applications. Can be self-hosted.

Proxy and dashboard for logging and analysing LLM requests.

Client libraries for LangSmith, LangChain's tracing and evaluation service.

The service itself is closed source.

Client library for the Weights & Biases experiment tracker.

The tracking server is a commercial product.

Per-second monitoring of servers, containers and applications.

Node-graph interface and backend for diffusion models.

In this workshop: Renders images on the home GPU.

Browser interface for Stable Diffusion with a large set of extensions.

The original Stable Diffusion release from CompVis.

No commits since June 2024.

Reference code for SDXL and Stability AI's later image and video models.

Inference code for the FLUX.1 image models.

The code is Apache-2.0. Each set of weights has its own licence. No commits since July 2025.

Video generation model.

The licence states that it does not apply in the EU, the UK and South Korea.

Smaller follow-up to HunyuanVideo.

The licence states that it does not apply in the EU, the UK and South Korea.

Image-to-video model built on HunyuanVideo.

The licence states that it does not apply in the EU, the UK and South Korea.

Video generation model.

Text-to-video and image-to-video diffusion models from Tencent AI Lab.

World models, datasets and tools for training robots and vehicles.

Platform for reproducible world-model research and evaluation.

Makes videos from React components.

Source-available. Free for individuals and small teams, paid for larger companies.

Generates a short video from a topic: script, footage, voice and subtitles.

Automated short-video engine that goes from a topic to a finished clip.

Formerly AIDC-AI/Pixelle-Video.

Image model that swaps FLUX's text encoder for the Qwen2-VL vision-language model.

No commits since November 2024.

Open voice model for long speech with several speakers.

Text-to-speech models with streaming output.

Speech recognition models for many languages, including singing.

Text-to-audio model for speech, music and sound effects.

No commits since August 2024.

Toolkit for audio, music and speech generation research.

List of open-source projects around Suno.

Unofficial JavaScript client for generating songs with Suno.

No commits since October 2024.

Workflow automation with a visual editor, code steps and over 400 integrations.

Fair-code: the source is public and self-hosting is free, reselling it is restricted.

Visual builder for agents and LLM workflows, with an API for each flow.

Workflow automation whose integrations also work as MCP servers.

Flow-based programming for wiring devices, APIs and services together.

Drag-and-drop builder for LLM agents.

Archived by its owner. Last commit August 2026.

Postgres platform with authentication, storage, realtime and generated APIs.

In-memory key-value store, forked from Redis.

Python client and CLI for downloading and publishing models and datasets.

Packages a machine-learning model as a container.

Customer relationship manager you can self-host.

Runs Linux containers in lightweight virtual machines on a Mac.

Renders Mermaid diagrams as themed SVG or ASCII art.

Desktop app for managing Markdown knowledge bases.

Obsidian's registry of community plugins and themes.

Obsidian itself is closed source.

Guides for rebuilding well-known tools from scratch: a database, git, a shell.

Tutorials grouped by language, each one building a complete project.

Step-by-step roadmaps for learning a role or a technology.

Formerly kamranahmedse/developer-roadmap.

Andrej Karpathy's course on building neural networks from first principles.

No commits since August 2024.

Andrej Karpathy's complete small ChatGPT pipeline, trainable for about $100.

Short, readable PyTorch implementation of GPT training.

No commits since August 2024.

Course on designing, building and deploying a machine-learning system.

Notebooks for Aurélien Géron's Hands-On Machine Learning, third edition.

Jake VanderPlas's book, complete, as Jupyter notebooks.

No commits since June 2024.

Summaries and resources for Chip Huyen's Designing Machine Learning Systems.

Research updates, courses, notebooks and interview material on generative AI.

The index of awesome lists, one per topic.

Python frameworks, libraries and tools, sorted by purpose.

Machine-learning libraries, sorted by language.

Reading path and resources for learning data science.

LLM papers, models, tools and courses.

No commits since July 2025.

Tools for deploying and operating LLMs.

Reading list for MLOps.

No commits since November 2024.

Open agent projects and frameworks.

Autonomous agents built on LLMs, with papers.

Agent frameworks and tools, kept for 2026.

Projects on LLMs, vision-language models and generated content.

No commits since August 2025.

Tools and resources for solo founders.

Tools for people building products alone.

No commits since December 2024.