OpenHands
Open-source autonomous software-development agent running inside sandboxed environments.
About
OpenHands is an open-source autonomous software-development agent that plans, writes code, runs terminal commands, and browses the web inside sandboxed Docker environments.
In 2026 the project reorganized: the main repository is now Agent Canvas, a self-hosted control center that can drive OpenHands as well as third-party agents (Claude Code, Codex, Gemini CLI) via the Agent Client Protocol, while the agent runtime moved to a separate Python SDK repository.
It is model-agnostic via a LiteLLM provider layer, MIT-licensed, and had roughly 91k GitHub stars as of October 2026. A managed OpenHands Cloud offering exists for teams that do not want to self-host.
What it can do
- Autonomous agent that writes code, runs shell commands, and browses the web end-to-end
- Sandboxed Docker execution isolates agent commands from the host machine
- Agent Canvas: self-hosted control center for conversations and scheduled/event-driven automations
- Model-agnostic via LiteLLM (100+ providers: Anthropic, OpenAI, Gemini, local models via Ollama, Bedrock, Vertex, and more)
- Runs third-party ACP-compatible agents (Claude Code, Codex, Gemini CLI, OpenCode) under one UI
- Software Agent SDK (Python) plus Agent Server with REST and WebSocket APIs
- Slack, GitHub, Linear, and Notion integrations for automations
- CLI and headless interfaces for scripting and CI use
Setup
- Install prerequisites: Node.js 24 or later,
uv, and Docker (Desktop on macOS/Windows, Engine or Desktop on Linux) for the sandboxed option. - Install the Agent Canvas launcher:
npm install -g @openhands/agent-canvas. - Start it with
agent-canvas— this starts the full local stack by default. Open http://localhost:8000 and start a conversation. - Alternative (Docker sandbox): create a host directory for your projects, set PROJECTS_PATH to it, and run the published agent-canvas image with your projects and
~/.openhandsstate mounted; the agent can then access only that directory. - Add your LLM provider API key in the UI settings (bring-your-own-model) and start a task, or define automations that run on a schedule or in response to webhooks.
Prerequisites
Node.js 24+, uv, Docker (for sandboxed mode); an LLM provider API key
Use cases
Works with
Trust score — 9/10
Very active (~91k stars, commits through Oct 2026), MIT, and the Docker-sandboxed execution model is a sound security posture for an autonomous agent. Caveats: the 2026 restructure (main repo is now "Agent Canvas"; agent runtime split into OpenHands/software-agent-sdk; legacy archived) makes docs navigation confusing, and published SWE-bench scores (68-72%) are reported by third parties, not independently verified by me. Metrics-v1 score 9/10 — adoption 8.4, maintenance 9.6, openness 10, manual 9.
★ 90,735 GitHub stars
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.4/10
- Maintenance · recency, releases, issue load
- 9.6/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · 90,736 GitHub stars (log-scaled vs catalog peers)
- · blended with ~18,659 monthly package downloads
- · last push 0d ago; 63 releases/12mo; 953 open issues
- · manual review: Docker-sandboxed posture sound; 2026 restructure confuses docs; SWE-bench scores third-party-reported
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.