OpenHands 1.0: an open, self-hostable coding agent at 68% SWE-bench
The agent rebuilt on an open SDK you can self-host, keep your keys, gate risky actions, and run on any model — from $0.20 a task. Devstral 24B hits 46.8%.
Copy markdown
Production-ready, and MIT
OpenHands 1.0 is a ground-up rebuild around its Software Agent SDK, scoring ~68% on SWE-bench Verified (500 real GitHub issues) and shipping under the MIT license with 86.9K stars. It's the open, model-agnostic answer to Devin you can actually read, fork, and run yourself.
Your machine, your keys
It runs in your own environment via Docker, and SANDBOX_VOLUMES mounts only the folders you choose — limiting the blast radius when the agent goes wrong. It's model-agnostic too: point it at Ollama, Bedrock, or a commercial API. Self-hosting runs roughly $0.20–$1.05 per resolved task at H100 rates.
The SDK is the real unlock
Four Python packages — openhands.sdk, .tools, .workspace, and .agent_server (REST + WebSocket) — let you build your own production agent instead of just running theirs. Paired with the open-weight Devstral 24B it hits 46.8% on SWE-bench Verified, matching Devin 2.0 with zero API spend.
Guardrails you can enforce
A SecurityAnalyzer rates every tool call low/medium/high risk, and the ConfirmRisky policy blocks high-risk actions until a human approves. Event-sourced history gives you full audit trails and deterministic replay — autonomy with a kill switch, not a black box.