I Made an Open-Source AI Fix My Code & Open a Pull Request (OpenHands)

An open-source AI software engineer read my broken repo, found the bug, fixed it, ran the tests, and opened a pull request — completely on its own. I just merged it. In this video I deploy OpenHands (formerly OpenDevin) on a Hostinger VPS, connect it to GitHub, and hand it a real bug to fix end-to-end.

What OpenHands is

OpenHands is a free, open-source, self-hosted autonomous coding agent. Not autocomplete, not a chat window — an agent that gets a goal ("fix issue #3"), then works the loop every autonomous agent runs:

Perceive — read the repo, the issue, the test output.
Plan — decide what to change.
Act — edit files, run commands.
Observe — check the result, and loop until the tests pass.

Everything runs inside a sandboxed container, so the agent can execute code and shell commands without touching the host system. It works with any LLM — I ran it with Claude via OpenRouter.

The Devin comparison

Devin made the "AI software engineer" pitch famous, but it's closed and paywalled. OpenHands gives you the same category of tool with the code on GitHub and the deployment on your own box. And the benchmark backs it up: 68.4% on SWE-bench Verified — top of the open-source pack. That's real-repo bug fixing, not toy puzzles.

Pros and cons

Pros

Free and open source, self-hosted — your code never leaves your server
Model-agnostic: bring Claude, GPT, or a local model
Full agent loop with test execution, not just patch suggestions
GitHub-native: reads issues, opens PRs

Cons

LLM API costs are on you, and agent loops burn tokens
Complex multi-file refactors can still wander
You babysit the first few runs before you trust it

The demo, end to end

The full walkthrough in the video:

Demo project — a small repo with tests: github.com/PromptEngineer48/openhands-demo
Plant a bug — break the code, watch tests fail, and file it as a GitHub issue like a normal team would.
Spin up the VPS — Hostinger KVM2 (hostinger.com/prompt, coupon PROMPT for 10% off), SSH in, install OpenHands.
Connect the pieces — the GUI, an OpenRouter key for the LLM, and a GitHub token for the repo.
Prompt it — point OpenHands at the issue and let go.

From there it's hands-off: the agent reads the issue, locates the bug, patches the code, runs the test suite, sees green, and opens a pull request with a description of the change. I review, merge, verify. The PR in my repo was written entirely by the agent.

Verdict

This is the part of AI coding that actually feels like a shift: not faster typing, but issue in, pull request out. OpenHands is the strongest open-source take on it right now — 68% SWE-bench, any LLM, running on a $10-class VPS you control. Set it up on a throwaway repo first, watch it work once, and you'll start filing issues differently.

Repo: github.com/All-Hands-AI/OpenHands · Docs: docs.openhands.dev