An autonomous agent with a sandbox, formerly OpenDevin
OpenHands is an agent that does software-engineering tasks end to end: it writes code, runs commands, and browses the web to get a job done with little or no hand-holding. The detail that separates it from an IDE assistant is the sandbox. Instead of editing files directly in your editor, OpenHands runs the agent inside an isolated runtime, which is what makes hands-off, multi-step work safer to let run. That design comes from its research roots, and it is worth knowing the lineage: the project started as OpenDevin, moved under the All-Hands-AI organization, and now lives at the OpenHands organization. If you find old references to OpenDevin or All-Hands-AI/OpenHands, they point here.
At 77,254 stars as of 2026-06 it is the largest of the open-source coding agents, and it pushes commits daily, so it is firmly active.
What it does
You give it a task, and the agent plans, edits files, executes shell commands, and reads their output, iterating until it believes the task is done or it needs you. Because it can run code in its sandbox, it can do things an editor assistant cannot safely attempt, like running a full test suite or spinning up a service to check behavior. It works with a wide range of models through standard provider APIs, including local ones, though local models come with a caveat covered below. Its standing on the SWE-bench benchmark, which measures resolving real GitHub issues, is one of the project’s headline claims and the reason it gets cited in agent research.
The four ways to run it
OpenHands is not a single binary. The README describes four surfaces, and choosing the right one saves a lot of confusion:
- SDK: a composable Python library that is the engine behind everything else, for defining agents in code.
- CLI: the easiest entry point, with an experience similar to Claude Code or Codex, powered by any LLM.
- Local GUI: a React app plus REST API for running agents on your laptop, similar in feel to Devin or Jules.
- Cloud: a hosted deployment you can try free with the Minimax model by signing in with GitHub or GitLab, with Slack, Jira, and Linear integrations and multi-user support.
Install and the Docker requirement
Start with the CLI if you just want to try it; the README points to the install command in the docs at docs.openhands.dev. The Local GUI is the heavier path: it runs the agent’s sandbox runtime in Docker, so you need Docker installed and configured, and the README’s setup guide is the source to follow rather than a command pasted from memory. The practical takeaway is that this is not a pip install and go tool for the full GUI experience; the sandbox is a real dependency, and that is the price of the isolation that makes the agent safe to run autonomously.
Where it fits, and where it doesn’t
Reach for OpenHands when you want genuinely hands-off work, like having an agent resolve an issue or refactor across files while you do something else, and when the sandbox isolation matters because you do not want an agent touching your live environment directly. The cloud and enterprise options make it the most plausible of these tools for a team that needs multi-user access and integrations.
It is the wrong tool for a quick inline edit, where its setup weight is pure overhead, and it is harder to adopt if you cannot or will not run Docker. The license is mostly MIT, but note the split: the enterprise/ directory is source-available, and running it for more than a month requires purchasing a license, so read that boundary if you are deploying the enterprise features.
OpenHands versus the lighter agents
| OpenHands | cline | Aider | Continue | |
|---|---|---|---|---|
| Stars | 77,254 | 63,353 | 46,274 | 33,704 |
| Shape | sandboxed autonomous agent | IDE agent with approval gates | terminal pair programmer | embedded assistant (read-only) |
| Runtime | Docker sandbox | your editor | your shell | your editor |
| Setup weight | heavy | light | light | light |
| Best at | hands-off multi-step tasks | reviewed edits in the IDE | git-native terminal work | inline help (frozen) |
Counts are from GitHub as of June 2026. cline keeps a human in the loop in your editor and is much lighter to set up. Aider is the terminal, git-native option. Continue was the embedded assistant and is now read-only. OpenHands trades setup weight for the ability to run autonomously in isolation, which is the whole point.
The star curve
The star curve has two acts: an early surge as OpenDevin when the autonomous-agent idea caught fire, then sustained growth after the rename and the move to a real platform with SDK, CLI, and cloud. The size of the count reflects both research interest and a genuine product, not a single viral week.
What the issue tracker warns you about
The recurring complaints cluster around the runtime. Local models can hit timeouts: the tracker has a long-standing issue where Ollama tasks are killed after a few minutes regardless of timeout settings, so treat local-model use as best for short tasks until that is resolved. Sandbox limitations show up too, like trouble opening browser preview links from inside the sandbox, and sessions that do not resume after a sandbox restart because the session id is not persisted. LLM configuration is another sharp edge: editing one model profile has been reported to overwrite others, and the web UI’s settings can overwrite a hand-edited settings.json, so decide on one place to manage configuration and stick to it. None of these block the core use case, but they explain why the smoothest experience is a frontier hosted model rather than a local one.
Related
For a lighter IDE agent with approval gates, see cline. For terminal, git-native work, see Aider. For the now read-only embedded assistant, see Continue. To keep long tasks organized across any agent, see Task Master. For what is trending, see LLM tooling, the daily digest, and the weekly report.
FAQ
Is OpenHands the same as OpenDevin? Yes. OpenDevin was the original name. The project moved under All-Hands-AI and now lives at the OpenHands organization, so old links and papers referring to OpenDevin or All-Hands-AI/OpenHands are the same project.
Do I need Docker to run OpenHands? For the Local GUI, yes. It runs the agent’s sandbox runtime in Docker, which is a real dependency. The CLI is the lighter way to start, and the cloud version needs nothing local at all.
Can OpenHands use a local model like Ollama? It can connect to local models, but the tracker has a long-standing timeout issue where Ollama tasks are killed after a few minutes, so local models are best for short tasks today. A frontier hosted model gives the most reliable experience.
OpenHands vs cline, which should I use? OpenHands runs autonomously in a Docker sandbox and is heavier to set up, which fits hands-off multi-step tasks. cline runs in your editor with approval gates and is much lighter. Choose by how much autonomy and isolation you want versus setup effort.
Is OpenHands free? The core is MIT-licensed and free. The cloud has a free tier with the Minimax model, and the enterprise/ directory is source-available, requiring a purchased license to run for more than a month.