Authority AI Editorial
OpenHands: an open-source coding agent
OpenHands is an open-source, self-hosted platform for running coding agents against a real development environment: a sandbox with a shell, a browser and a file system the agent can act on. It began in March 2024 as OpenDevin, a community response to the closed Devin agent from Cognition, and was renamed OpenHands around version 0.9.0 in August 2024. Its three co-founders, Graham Neubig, Robert Brennan and Xingyao Wang, also formed the company All Hands AI to back the open-source project and run a hosted cloud version. It is for developers who want to run an agent on their own infrastructure, keep code and credentials off a third-party server, or plug more than one agent backend into the same control surface.
What it does
OpenHands describes itself as a developer control center for coding agents and automations. By default it runs its own open-source agent, but it also supports third-party agents such as Claude Code and Codex through the Agent-Client Protocol, so you are not locked into one model or one agent implementation. Work happens inside a sandboxed runtime, most commonly a Docker container, so the agent can run shell commands, install dependencies and edit files without touching your host machine directly. On top of the core agent loop, OpenHands adds workflow automations that connect to tools like Slack, GitHub and Linear, so an agent run can be triggered by, or report back to, systems your team already uses, for tasks like decomposing a GitHub issue or generating a status report.
How you use it
The current documentation, at docs.openhands.dev (the older docs.all-hands.dev address now redirects there), organizes the project around a few pieces: Agent Canvas, the browser-based client and control center for agent conversations that can run locally or point at a remote backend; the Software Agent SDK, a Python library for building agents programmatically; and the Sandbox Server, a community-supported standalone API for managing sandboxes. For anyone who wants a managed option instead of self-hosting, OpenHands Cloud offers hosted access with integrations and access controls, and OpenHands Enterprise is a licensed self-hosting option for organizations. The project also notes that its code now spans more than one repository, with the Agent Server and SDK maintained separately in a software-agent-sdk repository rather than the main OpenHands repository. To get running, you can install the client with a global npm install, run it in Docker with your project directory mounted as a volume, or build from source; the newer stack expects Node.js 24 or later. An older CLI tool and a Local GUI still exist in the project's history but are called out in the docs as legacy and no longer actively developed, so new setups should start from Agent Canvas or the SDK rather than those paths.
What to know before adopting it
OpenHands is MIT licensed at the core project, though the documentation is explicit that each public repository in the OpenHands organization carries its own license, so check the specific repo you depend on rather than assuming one blanket license covers everything, especially if you look at the Enterprise offering. Because the default runtime is a sandboxed container, running it seriously means running Docker (or an equivalent) on whatever machine hosts the agent, which is a heavier local footprint than a lightweight CLI tool. There is no cost from the open-source project itself beyond your own model API usage and compute, but OpenHands Cloud and Enterprise are paid, commercially supported paths if you want managed hosting or organizational licensing instead of running the stack yourself.
Where it sits next to other agent tools
OpenHands is best understood as infrastructure and a control surface for coding agents rather than a single agent personality. That puts it in a different category from a framework like AutoGen, which is about how multiple agents talk to each other, and closer to being a home for whichever coding agent you choose to run, including agents you already use like Claude Code or Codex, inside a sandbox you control. If you already have a coding agent you like and mainly want a self-hosted place to run it safely with workflow hooks into Slack and GitHub, OpenHands is a direct fit. If you want the multi-role planning MetaGPT provides, or the multi-agent conversation patterns AutoGen provides, OpenHands is not that layer, though nothing stops you from running an agent built with either inside an OpenHands sandbox. Whatever agent ends up running inside it, the AI Ultimate Skill Collection packages 3,029 ready-made skills in 12 packs, written procedures your agent follows step by step.