An open source AI agent is a program you can read, build and change that gives a language model tools and runs it in a loop until a task is done: it browses, edits files, runs commands, calls APIs. The ones worth running in 2026 are goose for general work on the desktop, OpenCode for code in the terminal, Hermes Agent for an assistant that lives in Telegram, Slack or email, Agent Zero for an agent with its own Linux computer, and OpenHands when you want a dashboard that runs several agents at once. Below them sit the frameworks, LangGraph, CrewAI and Microsoft’s Agent Framework, which are not agents but the parts to build one.
The license is checked for every row, because “on GitHub” and “open source” stopped meaning the same thing: n8n, AutoGPT’s platform and Dify all publish their code under terms that are not open source, and Claude Code’s public repository reserves all rights.
Open source AI agents compared
| Agent | License | What it’s for | Runs where | Models | Maintained by |
|---|---|---|---|---|---|
| goose | Apache-2.0 | General work: research, writing, automation, code | Desktop app (macOS, Linux, Windows), CLI, API | 15+ providers; Ollama; Claude, ChatGPT and Gemini plans through ACP agents | Agentic AI Foundation at the Linux Foundation; started by Block |
| OpenCode | MIT | Coding | Terminal, desktop app (beta), IDE extension, web | 75+ providers; ChatGPT and Copilot sign-in; Ollama, LM Studio, llama.cpp | Anomaly Innovations |
| Hermes Agent | MIT | A personal assistant with memory that learns skills | Desktop (macOS, Windows), terminal on Linux, Telegram, Discord, Slack, WhatsApp, Signal, email | Any provider; Nous Portal with 200+ models | Nous Research |
| Agent Zero | MIT | An agent with a full Linux computer: terminal, browser, desktop apps | Docker, with a desktop launcher; web UI on localhost | Cloud providers, ChatGPT sign-in, local models | Agent Zero |
| OpenHands | MIT | Software development; Agent Canvas runs other agents too | Self-hosted web app, CLI, Docker, cloud | Any model; runs Claude Code, Codex, OpenCode, Pi and ACP agents | All Hands AI |
| Codex | Apache-2.0 | Coding | Terminal, IDE extension, ChatGPT desktop app, cloud | OpenAI with a ChatGPT plan or key; Ollama and LM Studio with --oss | OpenAI |
| Cline | Apache-2.0 | Coding in the editor | VS Code, JetBrains, CLI, desktop app | Many providers with your keys; ChatGPT sign-in | Cline Bot Inc. |
| Kilo Code | MIT | Coding in the editor, with a model gateway | VS Code, JetBrains, CLI, desktop app (beta), cloud | 500+ models through Kilo Gateway, your keys, Ollama, LM Studio | Anaconda, since July 2026 |
| Pi | MIT | A minimal coding harness you extend yourself; powers OpenClaw | Terminal, SDK | Many providers; llama.cpp built in | Mario Zechner |
| DeepSeek Harness | MIT | Coding; a developer preview | Local web UI, CLI, ACP | DeepSeek, plus Anthropic, OpenAI, Kimi, GLM and custom endpoints | DeepSeek |
| Aider | Apache-2.0 | Terminal pair programming with git commits per change | Terminal | Most cloud and local models | Paul Gauthier; last release February 2026 |
| Contenox | Apache-2.0 | Agents that ask first, under rules you can read | Desktop app (Windows first), contenox-runtime for scripts | ChatGPT plan, API keys for most providers, Ollama and vLLM; hosts Claude Code | Alexander Ertli |
Every row was checked against the project’s own repository or docs in November 2026. OpenHands is the most-starred on this list; goose is the one with a foundation behind it, which matters if you’re choosing for a company.
Agent or framework?
Half the “best open source AI agent” lists are really framework lists. The difference decides whether you have something to run tonight or something to build for a month.
| What you get | Examples | License | |
|---|---|---|---|
| Agent | A finished program: install it, give it a task | goose, OpenCode, Hermes, Agent Zero, OpenHands | See the table above |
| Framework | The loop, state and tool plumbing as a library; you write the agent | LangGraph, CrewAI, Microsoft Agent Framework, the OpenAI Agents SDK, the Claude Agent SDK | MIT (LangGraph, CrewAI, Agent Framework); SDKs under their vendors’ terms |
| Visual builder | A canvas where agents are workflows of boxes | n8n, Dify, Flowise | n8n: Sustainable Use License, source-available; Dify: Apache-2.0 with added conditions; Flowise: archived |
Two framework notes that change the choice: AutoGen is in maintenance mode, and Microsoft points new users to its Agent Framework. AutoGPT, the project that started the category in 2023, is now a platform under the PolyForm Shield license for its core, with MIT for the rest; it is not open source in the OSI sense.
If you want to build rather than run, the loop you’d be building is about twelve lines; what is an AI harness? shows it, and what the frameworks add on top.
Which open source AI agent should I pick?
- General tasks on your own computer, with a window: goose. Desktop app on all three systems, 70+ MCP extensions, and a foundation rather than a startup behind it.
- Coding in the terminal, any model: OpenCode. If you pay for ChatGPT, Codex; both are open source, and Codex has the only native Windows sandbox among them.
- An assistant you message from your phone: Hermes Agent. It keeps memories and builds its own skills across sessions, and runs on Telegram, Discord, Slack, WhatsApp, Signal and email.
- An agent that needs its own machine: Agent Zero. Each agent gets a Linux computer in Docker, with a browser and desktop apps you can watch or take over, and snapshots to roll back.
- Several agents on a team dashboard: OpenHands Agent Canvas. It runs its own agent and the others, self-hosted.
- Local models only: OpenCode, goose or Hermes, each with Ollama as a normal provider; or
ollama launch opencodeto skip the config. Run local LLMs for coding. - Something small enough to read in an evening: Pi. It has no permission prompts by design; run it in a container.
- Rules you can read before anything runs: Contenox. Allow, ask or block per tool and per command, in a desktop app.
What open source gets you, and what it doesn’t
It gets you:
- Code you can audit. What the agent sends, where, and which commands it may run are all in the source.
- No lock-in to the vendor. When Roo Code was archived in May 2026, its code stayed available to fork. When Kilo Code was sold to Anaconda in July 2026, the MIT license stayed.
- Model choice, usually. Most of these take any provider. Codex is the exception that proves it: open source, built around OpenAI.
It doesn’t get you:
- A free model. The agent costs nothing; the tokens don’t. One month of our own use on the DeepSeek API was 23,815 requests and 7.86 billion tokens, about 330,000 tokens per request. A local model is the only free one. AI coding costs.
- Privacy by default. Your files still go to whichever provider you pick. Only a local model keeps them on your machine.
- No telemetry by default. Open source means you can check, not that nothing is sent. Codex sends usage metrics unless
analytics.enabled = falseis set in its config. Read each tool’s settings. - Safe defaults. goose’s default mode never asks before acting. Pi has no permission prompts at all. DeepSeek Harness asks you to read a safety notice before running it. Open code and careful defaults are different things; AI agent permissions says what to set.
The source-available trap: n8n’s Sustainable Use License, AutoGPT’s PolyForm Shield and Dify’s extra conditions all restrict what you may do with the code, usually competing with the vendor or offering it as a service. Crush uses the Functional Source License, which only turns into MIT two years after each release. Claude Code’s repository is public, but its license reserves all rights. If the license matters to you, read the file, not the README badge.
Is Claude Code open source?
No. The repository is public; the code is used under Anthropic’s commercial terms. The closest open-source agents are Codex and OpenCode. See Claude Code alternative.
Which open-source agents are no longer maintained?
- AutoGen: maintenance mode; Microsoft Agent Framework is the successor.
- Flowise: repository archived, end of life.
- Roo Code: archived on 15 May 2026. Kilo Code and Cline are the usual next stops.
- Continue: read-only after its final 2.0.0 release.
- Rasa Open Source: maintenance mode; the commercial platform continues.
- Aider: still works, last release February 2026.
An unmaintained agent still runs. It won’t learn new models’ tool formats or get security fixes.
Contenox
Contenox is an open-source (Apache-2.0) desktop app, built Windows-first, for working with agents that ask first. You open a project folder, pick a provider, and a policy engine decides what each tool call may do, by rules you can read in Permissions, not by a model’s judgment of its own risk. It hosts Claude Code and other ACP agents under the same rules. Download it here.
Related guides
- Open source AI coding agents: the coding subset, with every license
- What is an AI harness?: what these programs are, under the hood
- Run local LLMs for coding: the only way to make the model free too
- AI agent permissions: what to allow, ask and block