OpenHands (formerly OpenDevin) is an open-source autonomous AI software agent developed by All Hands AI. It receives a task — a GitHub issue, a natural-language description, or a URL — and works through it inside a sandboxed Docker container that includes a full development environment: terminal, browser, code editor, and file system. The goal is a complete handoff from task description to pull request with minimal human intervention.
Key capabilities
Sandboxed Docker environment — Each task runs in an isolated Docker container. The agent installs dependencies, runs tests, browses the web, and edits files without touching the developer’s machine.
End-to-end task execution — OpenHands is designed for long-horizon tasks: reading a GitHub issue, exploring the codebase, writing a fix, running the test suite, iterating on failures, and opening a pull request with a description.
Browser and web access — The agent navigates web pages, reads documentation, searches for API references, and can interact with web UIs — making it less dependent on training data for up-to-date information.
Multi-provider LLM backend — OpenHands supports Claude, GPT-4o, Gemini, Llama, DeepSeek, and other models through a configurable backend. Teams can self-host with local models for privacy-sensitive workloads.
GitHub integration — Reads issues and pull requests, opens new PRs, and responds to review comments as sub-tasks.
Autonomy level
Level 4 (near-autonomous): OpenHands works end-to-end with infrequent human checkpoints. It makes architectural and implementation decisions independently and escalates to the human when genuinely blocked.
Strengths
- MIT license — fully open source, self-hostable
- Sandboxed execution provides a strong safety boundary
- Strong SWE-bench benchmark results among autonomous agents
- Model-agnostic with support for local LLMs
Limitations
- Web UI required; not a drop-in terminal tool for interactive development
- Docker dependency adds setup overhead
- Long-horizon tasks can accumulate errors that are hard to diagnose mid-run