Most “best open source AI agent” lists put a coding agent, a Python framework and a personal assistant in one ranking. That’s like ranking a hammer against a car. The useful question is what job you want done, then which open source project does that job.
This post sorts thirteen projects by job. Stars, licenses and last push dates come from the GitHub API on 2026-10-19. Disclosure: TODO for AI is ours. Its bridge and CLI are MIT, but the board and backend are hosted, so it is not ranked as a pure open source pick. Its section says what is open and what is not.
The short answer
| If you want | Pick | License |
|---|---|---|
| A personal assistant on your own machine, reachable from your chat apps | OpenClaw | MIT |
| The same, with an agent that writes its own skills | Hermes Agent | MIT |
| A free desktop app like Claude Cowork | OpenWork or Eigent | MIT desktop app / Apache-2.0 |
| A coding agent in the terminal | OpenCode | MIT |
| Coding agents running on your servers for a team | OpenHands | MIT |
| An agent that drives a web browser | browser-use | MIT |
| Manus-style research that runs fully local | AgenticSeek | GPL-3.0 |
| To build your own multi-agent system | CrewAI or LangGraph | MIT |
| A team board, with the agent running on your own machines | TODO for AI | MIT bridge and CLI, hosted board |
Personal assistants that run on your hardware
OpenClaw (392k stars, MIT). An assistant that runs on your computer and answers in the channels you already use: WhatsApp, Telegram, Slack, Discord, iMessage, Teams and 20+ more. State, memory and credentials stay on your machine. Models and agent harnesses (Claude, Codex, local models) are plugins. It is the most-starred project here by a wide margin. The catch is the same as every self-hosted agent: you run it, update it, and decide what it may touch.
Hermes Agent (255k stars, MIT). Nous Research’s agent with a learning loop: it writes skills from what it did and improves them as it uses them. Any model, including your own endpoint. One-line installers for Linux, macOS, native Windows and Android through Termux. Pick it over OpenClaw if you care more about the agent getting better at your tasks than about how many chat apps it can answer in.
Desktop apps like Claude Cowork
OpenWork (24k stars). Its tagline is “the open-source alternative to Claude Cowork”. A desktop app for macOS, Windows and Linux that works on your local files, built on OpenCode. Any model, including Ollama. Skills, Claude-compatible plugins and MCP servers carry over from Cowork. The README says the desktop app is MIT, but GitHub cannot detect one license for the whole repo, so check the folder you plan to ship. The team control plane can be self-hosted.
Eigent (16k stars, Apache-2.0). Calls itself “the open source Cowork desktop”. It runs local and free, and is aimed at the same job as OpenWork. Its Apache license is the cleaner one if you plan to embed it in a commercial product.
Coding agents
OpenCode (217k stars, MIT). The open source coding agent for the terminal, and the base OpenWork is built on. Works with any model. The default answer to “open source alternative to Claude Code”.
OpenHands (91k stars, MIT). It has moved from being one coding agent to a control center for many: Agent Canvas runs OpenHands, Claude Code, Codex, Gemini CLI or OpenCode on your laptop, in Docker, on VMs or on your company’s servers, and connects to Slack, GitHub and Linear. Pick it when a team wants coding agents that keep running after laptops are closed.
Browser agents
browser-use (118k stars, MIT). A Python library that lets a model drive a browser. It is a library, not an app: you write the script. The company also sells a hosted cloud with stealth browsers, but the library is MIT. It is the default building block for “AI that clicks through websites”.
Manus-style research, fully local
AgenticSeek (28k stars, GPL-3.0). Pitched as “fully local Manus AI”: it thinks, browses and writes code on your machine with local models and no API bills. The GPL matters if you plan to ship it inside closed software. Local models also mean your hardware sets the quality bar.
Frameworks for building your own
CrewAI (60k stars, MIT). Role-based multi-agent framework: define agents with roles and goals, give them tasks, let them work as a crew. The fastest way from idea to a working multi-agent prototype.
LangGraph (44k stars, MIT). Agents as graphs with explicit state. More code than CrewAI, more control over every step. Pick it when the agent has to pause, resume, retry or wait for a human.
smolagents (30k stars, Apache-2.0). Hugging Face’s barebones library for agents that act by writing code. Small enough to read in an afternoon.
AutoGen is archived. Microsoft no longer maintains AutoGen and points new users to Microsoft Agent Framework. Many lists still recommend it. Don’t start a new project on it.
AutoGPT (188k stars) is now a platform with a visual agent builder. GitHub cannot detect one license for the repo, so read the license files before you build on it.
TODO for AI: open source runtime, hosted team board
TODO for AI is a shared to-do list where each item is done by an agent. The parts that touch your machines are open source:
- Bridge (MIT). A native C binary, under 260 KB, that gives the agent a shell on your PC, server or cloud VM over a Noise-encrypted connection. No OpenSSL, no runtime to install.
- CLI (MIT).
npm install -g @todoforai/cli, thentfa-cli "Fix the login bug"runs an agent on your machine with any supported model.
The board, the agents’ coordination and the model routing are hosted, not open source. So you cannot self-host the whole thing the way you can OpenClaw or Hermes Agent. In exchange, a team gets one board where people and agents share the same tasks, a Chrome extension that works in your logged-in browser, a cloud VM that keeps running when your laptop is closed, and Claude, GPT, Gemini and Grok from one budget. Pricing · How it compares.
Side by side
| Job | Runs where | Models | License | Stars | |
|---|---|---|---|---|---|
| OpenClaw | Personal assistant | Your machine | Any, as plugins | MIT | 392k |
| Hermes Agent | Personal assistant that learns | Your machine or VPS | Any | MIT | 255k |
| OpenCode | Coding | Your terminal | Any | MIT | 217k |
| AutoGPT | Agent builder platform | Self-hosted or cloud | Any | Mixed, check files | 188k |
| browser-use | Browser automation library | Your code | Any | MIT | 118k |
| OpenHands | Coding agents for a team | Laptop, Docker, VMs | Any | MIT | 91k |
| CrewAI | Multi-agent framework | Your code | Any | MIT | 60k |
| LangGraph | Stateful agent framework | Your code | Any | MIT | 44k |
| smolagents | Minimal agent library | Your code | Any | Apache-2.0 | 30k |
| AgenticSeek | Local Manus-style research | Your machine | Local | GPL-3.0 | 28k |
| OpenWork | Cowork-style desktop | Your machine | Any, incl. Ollama | MIT desktop app | 24k |
| Eigent | Cowork-style desktop | Your machine | Any | Apache-2.0 | 16k |
| TODO for AI | Team board, agents on your machines | Your PC, servers, cloud VM | Claude, GPT, Gemini, Grok | MIT bridge and CLI, hosted board | new |
How to choose
Do you want an app or a library? If you want to use an agent today, pick from OpenClaw, Hermes Agent, OpenWork, Eigent, OpenCode or OpenHands. If you want to build one into your product, pick from CrewAI, LangGraph, smolagents or browser-use.
Is it you or your team? Most of these are built for one person on one machine. OpenHands is the exception for coding. For team work that isn’t code, OpenWork’s shared skills and TODO for AI’s shared board are the two options here.
Does the license matter? MIT and Apache-2.0 let you do almost anything. GPL-3.0 (AgenticSeek) requires you to open your changes if you distribute it. Where GitHub shows no single license (OpenWork, AutoGPT), read the files before you build on it.
For closed agents compared on the same jobs, see Best Claude Cowork alternatives and the alternatives table.
Stars, licenses and maintenance status were checked on the GitHub API on 2026-10-19.