Orbi: take GitHub issues through AI implementation and review
Orbi coordinates an AI development workflow around GitHub issues. It is designed to take a specified task through implementation, review and the repository's merge and release process, with the public GitHub record serving as evidence of the work.
An issue becomes a delivery workflow
The documented flow begins with an issue labeled ai-ready. Orbi implements it in an isolated worktree, then uses a separate session to review the change against the issue's acceptance criteria. That reviewer can make repairs and rerun the test suite.
The product emphasizes tying the review result to the exact commit that merges. Its recovery approach uses the issue, pull request, labels, journal and run identity to reconstruct interrupted work. These are useful features for teams who want an agent to handle a bounded backlog item rather than assist only while someone is editing.
Repository rules remain important
Orbi's FAQ says normal pull requests, branch protection, required checks and human approvals still apply. That gives a team familiar places to inspect and constrain a delivery. The value of a separate AI review session still depends on the quality of the acceptance criteria and tests; it does not establish that a change is safe merely because another agent accepted it.
A practical pilot should use a small repository task with a clear expected outcome. Inspect the implementation, the review's repairs and the final test results before broadening the scope.
Self-hosted and cloud options
The core is open source under AGPL-3.0, and the site offers self-hosting with a model supplied through the user's own configuration. The managed cloud option requires GitHub authorization and app installation and includes model usage in its plans.
Check the current setup guide for platform and dependency requirements. Self-hosting determines where the runner operates, while the chosen model endpoint and repository access determine additional data flows. Public delivery examples demonstrate specific projects, rather than a guaranteed unattended success rate for a different codebase.