OpenHands

What it is

OpenHands presents itself as the open platform for agentic software development. Its site offers a download link and a chat interface that asks the user what they want to build ^[1]. The premise is easy to state: rather than only suggesting lines of code inside an editor, an agent is given a goal and a working environment, and is expected to carry the task through — reading a repository, running commands, checking the output and producing a change.

From OpenDevin to OpenHands

The project was formerly known as OpenDevin ^[2]. Its paper frames the aim as letting agents interact with the world the way a human developer does: by writing code, using a command line and browsing the web ^[2]. The paper was submitted to arXiv on 23 July 2024 and was accepted at ICLR 2025 ^[2]. OpenHands is released under the permissive MIT license and is described as a community project spanning academia and industry ^[2]. That breadth is visible in the numbers the paper reports: more than 2.1K contributions from over 188 contributors ^[2].

How it works and how it is measured

The platform supports sandboxed environments for code execution, coordination between multiple agents, and the incorporation of evaluation benchmarks ^[2]. Sandboxing matters because an agent that writes and runs code needs room to fail: it can install dependencies, run tests and hit errors without touching the machine the user actually works on. Multi-agent coordination addresses tasks that are easier to split than to do in one pass. For evaluation, the authors test agents on 15 challenging tasks, including SWE-bench for software engineering and WebArena for web browsing ^[2]. Putting web browsing next to the command line reflects a simple observation about real work: much of it consists of reading documentation, following error messages and looking things up, not only writing functions.

The product surface

On the project's site, a workspace organises work as conversations tied to a repository, alongside automations such as a pull request review or a nightly security pass ^[1]. The same agent can therefore be pointed at a task by a person in a chat, or run repeatedly on a schedule against the same repository, which turns a one-off session into something closer to a standing part of a development workflow.

Components

The documentation describes Agent Canvas as the open-source browser client and control centre, which connects to one or more Agent Server backends ^[3]. The Software Agent SDK is a composable Python library, and the same repository exposes Agent Server through REST and WebSocket APIs ^[3]. OpenHands Cloud is the managed commercial service for running the platform without operating your own backend and sandbox infrastructure ^[3]. The documentation also notes that the legacy OpenHands CLI is no longer actively maintained and that the legacy local GUI launched with openhands serve is deprecated ^[3]. Read together, these pieces describe a layered system: a browser client for people, a Python library for building agents, a server for running them, and a hosted option for teams that would rather not run the machinery themselves.

Where it sits

OpenHands is unusual in being both a research artefact and a working open-source project. The paper supplies the design and the benchmark results, while the MIT license and the contributor base make it something people can download, modify and self-host; the cloud service covers those who would rather not maintain infrastructure. For anyone learning about AI agents, it is a concrete example of what the word implies in practice — not only conversation, but a model wired into a sandbox, a command line, a browser and version control.