OpenHands

OpenHands (formerly OpenDevin) is an open-source platform for AI-powered software engineering agents developed by All Hands AI. It enables AI agents — backed by models like Claude, GPT-4o, or Llama — to autonomously complete software engineering tasks: writing code, fixing bugs, running tests, navigating documentation, and deploying changes.

**Core approach:** OpenHands operates through a sandboxed environment where the agent has access to a code editor, terminal, and web browser. Given a task description ("add a login form to this Flask app with email validation and session management"), the agent plans the implementation, writes the code, runs tests, and iterates on failures — without requiring step-by-step human supervision.

**Agent capabilities:** - Full code editor access: read, write, create files - Terminal: run any shell command, install packages, execute scripts - Web browser: look up documentation, check GitHub issues, search Stack Overflow - Multi-file understanding: work across a codebase with context about the full project structure

**Supported models:** OpenHands is model-agnostic. Users configure it with their API keys for Claude Sonnet (generally recommended for coding tasks), GPT-4o, or open models via Ollama. Claude Sonnet 3.5/3.7 is frequently cited in community benchmarks as the most effective model for complex multi-file engineering tasks.

**SWE-bench performance:** OpenHands achieved strong results on SWE-bench (a benchmark for resolving real GitHub issues), placing it among the top-performing open-source agent systems for software engineering.

**Deployment:** Runs locally via Docker (one-line setup command). A hosted cloud version at app.all-hands.dev is in development for users who don't want to self-host.

**Use cases:** Bug fixing from issue descriptions, feature implementation, code refactoring, test generation, and codebase exploration/explanation. Works best with clearly specified tasks and well-structured codebases.

**Who it's for:** Developers who want to explore AI-assisted coding beyond autocomplete — where the agent can work on a full task rather than completing a single function. Also researchers studying AI agent behavior and capabilities.