How does OpenHands, the open-source AI that writes and fixes its own code, work?

OpenHands is an open-source AI coding agent that takes a task you give it, plans on its own, writes code, runs it, and when errors appear keeps fixing them in a loop until a chosen test passes. The software itself is free under the MIT license, it has passed 70,000 GitHub stars, and on the standard SWE-bench benchmark it resolves more than half of real GitHub issues, results compared with top commercial tools.

In one line: a free open-source (MIT) coding agent that writes, runs, and debugs its own code, runs on your own machine, and lets you plug in whichever LLM you want.

What exactly does OpenHands do?#

OpenHands is an 'AI agent' that is specialized in building software. An agent is an AI that goes beyond a chatbot answering questions: give it a goal and it breaks the work into steps and actually gets it done, like a tireless junior developer. Hand it a task and it (1) plans, (2) writes code, (3) runs it in a terminal, and (4) reads the logs and fixes errors by itself, then runs again. The key is that it repeats steps 2 through 4 in a loop until a test passes.

All of this happens inside a sandbox, a temporary run box walled off from the outside system. Because the agent really executes code, it is kept in an isolated room so a mistake cannot touch your real system, and a Docker container usually plays that room.

A common beginner misconception to clear up: OpenHands itself is not the 'brain.' The actual reasoning is done by the LLM you attach (a large language model, the chatbot's brain, such as Claude, GPT, or a local model on your own computer), while OpenHands is the harness that gives that brain 'hands and feet': a terminal, files, and a browser. So performance depends heavily on the model you connect; a weak model gives weak results. For reference, it was formerly called OpenDevin.

Why does this tool matter in 2026?#

The core shift of 2026 AI is packed into this one tool: moving from 'AI that answers' to 'AI that does the work' (agents). A chatbot only tells you "fix it this way," while an agent applies that fix itself and runs the tests to confirm the result.

Being open-source gives two concrete advantages. First, you are not locked to one vendor and can pick whichever LLM to connect: a strong cloud model like Claude or GPT if budget allows, or a local model if privacy matters. Second, its performance is publicly verified. On the standard SWE-bench benchmark (which has AI fix real bug issues from GitHub repositories and grades it with the original tests), it resolves more than half (about 53%) of real issues. But that number shifts a lot with the LLM attached, so read it as "OpenHands with a good model = 53%," not "OpenHands = 53%."

What is it good at, and still weak at?#

Choosing what to hand the agent is half the skill. Separate the tasks that fit from the ones it still fumbles.

The pattern is clear: OpenHands is stronger the clearer the finish line (the test) is. Give it "done when this test turns green" and it loops tirelessly until it gets there. Conversely, when the bar is vague like "refactor this nicely," the agent does not know where to stop and wanders or drifts off course. So the real skill lies less in coding and more in writing the bar that says 'when this test passes, you are done.'

What do you need to try it?#

The setup is simple. Install Docker (a tool that runs programs in isolation), plug in a model API key you already use, and it runs fairly quickly. The install method changes fast, so always check the official docs. Once running, a browser view lets you watch the agent plan and type code in real time, and you can stop and correct it mid-course if it drifts.

Three cautions. First, if you use an external model API like Claude or GPT, part of your code is sent to the provider, so for sensitive code use a local model together with a proper sandbox setup. Second, OpenHands actually executes the code it writes, so hardening the sandbox is entirely your responsibility; loose isolation is itself a security risk. Third, 'tests pass' does not equal 'correct.' A test can miss cases, so human review is still needed.

For the bigger picture, 2026 is the year AI moves from tool to coworker. OpenHands is less an autocomplete assistant and more an 'autonomous software engineer graded by tests.' The person who chooses what to delegate and writes a clear finish line decides how good this tool turns out to be. That is the shift Hax is here to watch.

Note: OpenHands is under active development, so the install steps, default model, and SWE-bench score change with each release. Verify the latest numbers and setup against the official GitHub repository and the SWE-bench leaderboard. Star counts and scores are estimates as of July 2026.