What Is OpenHands? The Open-Source Coding Agent
OpenHands is an open-source platform for AI coding agents, and it began in March 2024 as OpenDevin. Its agents write code and run commands inside a sandbox. In 2026 the project looks different from the version most tutorials describe. The old Docker GUI is marked legacy and the standalone CLI is unmaintained. The README now leads with a product called Agent Canvas. If you are evaluating OpenHands now, this guide shows how to run it today and what it costs. It ends with how OpenHands carries project knowledge between sessions.
What is OpenHands, in one sentence?
OpenHands is an MIT-licensed platform from All Hands AI for running software engineering agents that you can self-host or use in its cloud.
The OpenHands website calls it "an open source platform for software engineering agents." The GitHub repository had 92.1k stars on October 28, 2026. Its latest release was v1.34.0, published October 27, 2026. The LICENSE file is plain MIT.
The project started as OpenDevin on March 12, 2024, according to the team's one-year retrospective. Its founders formed All Hands AI, and the project was renamed. A September 2024 press release announced a $5M seed round led by Menlo Ventures. It named Robert Brennan, Xingyao Wang and Graham Neubig as founders.
How does OpenHands differ from Cline?
Cline is an agent you drive from your editor or terminal. OpenHands runs each agent in an isolated container and can host other agents next to its own.
| OpenHands | Cline | |
|---|---|---|
| License | MIT | Apache-2.0 |
| Where the agent runs | A Docker sandbox per session, or a remote workspace | Your editor or terminal |
| Models | Any model LiteLLM supports, with your key | Many providers, with your key |
| Project instructions | AGENTS.md plus skills in .agents/skills/ | .clinerules and AGENTS.md |
| Hosted option | OpenHands Cloud | No |
Sources: the OpenHands docs with its LLM page and skills page. Cline details are in our Cline guide.
The README describes Agent Canvas as a "self-hosted developer control center for coding agents and automations." It runs the OpenHands agent by default. It can also launch Claude Code, Codex and Gemini CLI.
How does OpenHands work?
The design comes from the team's paper, OpenHands: An Open Platform for AI Software Developers as Generalist Agents. It was accepted at ICLR 2025 and lists 24 authors. The agent and its environment talk through an event stream. Every action the agent takes and every observation it gets back is an event. The default agent was CodeActAgent, which acts by writing and running code.
Each session gets "a securely isolated docker container sandbox," per the paper. The agent edits files and runs shell commands inside that container when you use the default Docker setup.
The current version is built on the Software Agent SDK. Its repository calls it "a clean, modular SDK for building AI agents with OpenHands V1." A second paper, accepted at MLSys 2026, describes the SDK (arXiv 2511.03690). The architecture docs list several workspace types. DockerWorkspace runs the agent server in a container. LocalWorkspace runs on your machine. RemoteAPIWorkspace connects to a remote server.
For models, the LLM docs say OpenHands "can connect to any LLM supported by LiteLLM." The same page warns that it needs a powerful model. You can also point it at local servers such as Ollama or vLLM.
How do you run OpenHands today?
The docs home lists the current products:
- Agent Canvas, the self-hosted control center, free under MIT with your own LLM key
- OpenHands Cloud, the hosted version, where you log in with a Git hosting account
- the Software Agent SDK and Agent Server, for building your own agents
- Enterprise, for licensed self-hosting with commercial features
Two older paths are fading. The docs label the Docker-based local GUI as "Legacy Local GUI." The separate OpenHands-CLI repository says "This project is no longer actively maintained" and points users to Agent Canvas.
There is also a GitHub workflow for automated pull request review. The PR review guide explains the trigger. You add a review-this label to the pull request, or request openhands-agent as a reviewer. It needs an LLM API key stored as a repository secret.
What does OpenHands cost?
The pricing page lists three tiers when we checked on October 28, 2026:
| Plan | Price | What you get |
|---|---|---|
| Agent Canvas | Free | MIT license, self-hosted, bring your own LLM key |
| Teams | $500 a month for up to 10 users, then $50 per extra seat | 5,000 sessions a month, 30-day trial |
| Enterprise | Custom | Licensed self-hosting and commercial features |
Model costs are separate in every tier. The pricing page does not list a per-use price or free credits for OpenHands Cloud, so check your account before you plan a budget around it.
How does OpenHands remember a project?
OpenHands has two kinds of memory. One manages a single long session. The other carries project knowledge into new sessions.
Inside a session, the condenser keeps the context small. The condenser guide says it will "manage agent memory by condensing conversation history to save tokens." The default summarizes older events with an LLM and keeps the first few events intact. The same guide claims this cuts per-turn API costs by up to 2x, a vendor figure we could not check.
Across sessions, OpenHands uses files in your repository. The skills docs describe four kinds:
- AGENTS.md at the repository root, loaded as always-on context
- keyword-triggered skills that load when a word appears in your message
- path-triggered rules that apply to matching files
- on-demand skills, each a SKILL.md in its own folder under .agents/skills/
User-level skills go in ~/.agents/skills/. OpenHands also reads CLAUDE.md and GEMINI.md. The older .openhands/microagents/ folder still works, but the docs say to use .agents/skills/ for new skills.
Skills change only when a person, or the agent on request, edits them.
How good is OpenHands on benchmarks?
All Hands published a result in April 2025: 60.6% on SWE-bench Verified with one attempt, and 66.4% with five attempts plus a critic model. That comes from the company's own blog post and is a vendor claim.
The SWE-bench leaderboard data marks some entries as checked by the SWE-bench team. OpenHands with GPT-5 scored 71.8% on SWE-bench Verified in an August 2025 submission. OpenHands with Claude 4 Sonnet scored 70.4% in May 2025. These scores depend as much on the model as on the agent. We found no newer Verified result from All Hands, so treat these numbers as a 2025 snapshot.
The bottom line
Use OpenHands if you want an open-source agent that runs in a sandbox and can be self-hosted under MIT. Start with Agent Canvas and your own model key. Skip the legacy Docker GUI and the old CLI. Put project rules in AGENTS.md and add skills under .agents/skills/ as patterns repeat.
How do you share coding agent memory across OpenHands and other agents?
OpenHands reads AGENTS.md and skills from your repository. Those files only change when someone edits them. A fix the agent discovered on Tuesday is gone on Wednesday unless a person wrote it down. Agent Canvas can run Claude Code and Codex next to OpenHands, and each of those agents keeps its own context.
memU keeps what agents learn as a shared Markdown wiki. A scheduled background task slices new session logs into jobs. The agent distills each job into memory and skill Markdown, then commits it. memU's MemoryService makes no LLM or chat calls. It stores and embeds the Markdown the agent prepared, then retrieves it. A standing instruction in AGENTS.md or CLAUDE.md tells the agent to retrieve before answering. All hosts share one backend, so what Codex learned, Claude Code can retrieve.
OpenHands is not on memU's named host list. A generic adapter detects other agents, and the support matrix shows what works today.
To try the hosted version, which is free and syncs across devices, get an API key from memu.so and send your agent one message:
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU
memU is open source under Apache-2.0. The self-hosted version keeps data on one device and needs an embedding key.
FAQ
Is OpenHands the same as OpenDevin?
Yes. The project started as OpenDevin in March 2024 and was renamed OpenHands after its founders formed All Hands AI.
Is OpenHands free?
The software is free under the MIT license. Agent Canvas is free to self-host with your own LLM key. The Teams plan costs $500 a month for up to 10 users, and model usage is billed separately.
What models does OpenHands support?
Any model that LiteLLM supports, including hosted APIs and local servers such as Ollama. The docs warn that weak models perform poorly on real coding tasks.
Where does OpenHands read project instructions from?
From AGENTS.md at the repository root and from skills in .agents/skills/. It also reads CLAUDE.md and GEMINI.md, and the legacy .openhands/microagents/ folder still works.
Is the OpenHands CLI still maintained?
No. The OpenHands-CLI repository says it is no longer actively maintained and points users to Agent Canvas.