OpenHands

Overview

OpenHands is an open-source platform for autonomous software-development agents, built by All Hands AI. The project started in March 2024 as OpenDevin — a community response to the closed Devin demo — and was renamed OpenHands later that year. It is one of the most-starred AI agent repositories on GitHub, at roughly 87,000 stars and 11,000 forks as of September 2026, with 500+ contributors and, per the company, more than 9 million open-source downloads.

The technical premise is that a coding agent needs a real computer, not a chat box. Each OpenHands session runs in a sandboxed container where the agent can edit files, execute shell commands, install dependencies, run tests and browse the web, iterating on its own errors until a task is done. Crucially, the agent is model-agnostic: the same loop runs on Claude, GPT, Gemini or a locally hosted open-weight model, so teams choose their own cost/capability trade-off instead of inheriting a vendor’s.

All Hands AI, founded by Robert Brennan (CEO), Xingyao Wang and Carnegie Mellon’s Graham Neubig, raised a $5M seed led by Menlo Ventures in September 2024 and an $18.8M Series A in November 2025, bringing total funding to roughly $23.8M. Commercially, the model is the familiar open-core one: the agent itself is MIT-licensed and free forever, a hosted Individual tier passes LLM usage through at cost, and revenue comes from Enterprise deployments.

The trade-off is friction. Standing up the local version means Docker, a runtime image and API keys, and the free hosted tier’s 10-conversations-per-day cap makes it a trial rather than a daily driver. Like most long-horizon agents, it can also loop on ambiguous tasks and burn tokens without producing a usable diff. Teams that want to open a tab and start coding will find Cursor or Claude Code faster to adopt; teams that need to inspect, modify and self-host the agent itself have very few alternatives this mature.

Key Benefits

• No lock-in, by license: MIT licensing plus self-hosting means the agent, the runtime and the model are all replaceable — an unusual position among production-grade coding agents.
• Real execution environment: The agent runs commands and tests inside a sandbox, so it verifies its own work rather than emitting code that has never been executed.
• Model portability: Swap between frontier APIs and local open-weight models per task without changing workflow or tooling.
• Transparent economics: The hosted tier bills LLM usage at provider cost with no markup, making spend predictable and auditable.
• Deep community: A large contributor base means fast fixes, many integrations and a broad body of public configuration knowledge.

Use Cases

1. Issue-to-PR automation — Point the agent at a GitHub or GitLab issue and let it branch, implement, run the test suite and open a pull request for human review.
2. Regulated and air-gapped development — Self-host into a private VPC with a local model so no source code leaves the network, which is why security-conscious teams pick it over SaaS agents.
3. Legacy maintenance and migrations — Run repetitive, well-specified refactors — dependency bumps, framework migrations, test backfills — across a large codebase in parallel sessions.
4. Agent research and customization — Fork the agent loop to benchmark prompting strategies or scaffolding changes, a common use in academic SWE-bench work.