Orbi

Self-hosted coding factory that turns labeled GitHub Issues into independently reviewed, merged PRs and releases.

Open source alternative to:

Orbi is a self-hosted software delivery system for GitHub repositories. You write the goal and acceptance criteria in a GitHub Issue and add the ai-ready label; Orbi claims it, creates an isolated git worktree from a frozen base SHA, and hands the task to the Pi coding agent, which plans, implements, tests and commits. A runner then pushes the branch and opens a pull request whose body links back to the issue.

After the PR exists, a second, independent agent session reviews the exact base and head commits and fixes its own findings in the same session. Only the reviewed head can pass the merge gate, and Orbi never pushes or merges protected branches. If a failure is recoverable, the next tick resumes the same run and PR; if it is not, the issue is marked ai-blocked and waits for a human. Releases are their own labelled task that runs release gates and publishes the tag and notes.

What makes the design unusual is that GitHub Issues are the only state store: there is no database, queue or long-running daemon. A systemd user timer on Linux or a launchd agent on macOS fires a tick every five minutes, each tick performs at most one delivery step, and every log line and progress comment carries the same run_id so the whole timeline can be reconstructed. That replaces paid autonomous engineering services such as Devin, Codex cloud tasks or the GitHub Copilot coding agent with something you run on your own machine against your own repositories, for free, under the AGPL-3.0 (or the optional Sustainable Use License).

Frequently Asked Questions

What do I need to run Orbi on my own machine?

A GitHub repository, git, the GitHub CLI at version 2.94 or newer, uv (which provisions Python 3.14), and Node 22.19 or newer for the Pi coding agent CLI, which the installer also sets up. You also need a user scheduler session: systemd on Linux, or launchd on macOS. The macOS path exists but is not yet verified on real hardware according to the documentation.

Which AI models or providers can Orbi use?

Orbi does not talk to models directly. It drives the Pi coding agent CLI, so any provider already configured in Pi works, and a pi_providers file can point it at an OpenAI-compatible endpoint such as a local llama.cpp server. One real pi --print call is the check that a provider is usable before setup.

Does Orbi need a database, queue, or always-running server?

No. GitHub Issues are the only state store, and progress is recorded in issue and PR comments plus journal logs. A scheduler timer fires a tick every five minutes, and each tick does at most one delivery step, so nothing polls or supervises in between. Up to five runner instances can run concurrently, depending on max_concurrency.

Can I run Orbi from a container?

Yes. Published images are available on GitHub Container Registry as ghcr.io/orbi-build/orbi and on Docker Hub as orbibuild/orbi. Recent releases read ORBI_PI_ environment variables inside the container so the model provider can be configured there.

What happens when the agent cannot finish a task?

Failures are classed as recoverable or not. Recoverable ones, such as a failed test run or a stale base, return to the same run, branch, worktree and pull request on the next tick. Unrecoverable ones mark the issue ai-blocked with the full scene in a comment, and a human decides what happens next.

Visit Orbi