Orbi

Purpose Orbi is positioned as an open-source (AGPL-3.0) fully automated software factory for the agent era, with the core goal of driving GitHub Issues directly to merged PRs and formal releases. It is officially described as "File an Issue. Get a release.": file an issue, get a release. It primarily addresses the problem of having AI complete the entire delivery chain from coding to review, merge, and release in an unattended manner.

Main capabilities

• Issues as executable units: Orbi does not manage individual issues in isolation; instead, it uses Epics and milestones to grasp scope and dependency relationships to arrange the order of work.
• Isolated worktree coding: after taking on an issue, Orbi writes code in an isolated worktree.
• Independent review and fixes: another independent AI session reviews the specific diff; this reviewer can modify the code, rerun the test suite, and make a judgment on the fixed commit, and only the commit that passes review is allowed to merge.
• Merge and release: only commits that pass review are merged; Orbi then tags the release; a release issue is used to freeze the result as a version.
• Verifiable delivery records: issues, PRs, tags, logs, and run IDs are all kept in GitHub; after a restart, the same delivery (same run, branch, worktree, and PR) can be reconstructed from these records, and neither a killed process nor a failed test will produce a "second account."
• Delivery evidence: including fixable independent review, merge evidence bound to a single SHA, and a recovery mechanism that requires no second narrative.
• Does not replace final human gatekeeping: work results land as normal GitHub Pull Requests, and branch protection, required checks, and human approval apply as usual.

How to use

• Self-hosting: requires Python 3.14, git, gh, systemd, and a model endpoint; running orbi doctor can point out missing dependencies. Self-hosting requires Git, a POSIX shell, and model API access that you have already paid for. Orbi runs on Linux and macOS and can run on your own machine or server; it does not require a GPU—inference is left to the model, which can be a hosted API.
• Model access: when self-hosting, bring your own model; Orbi drives the agent through your credentials and supports any OpenAI-compatible API Or locally hosted models, after which the model can be swapped without changing the workflow. Model fees are paid to the model provider at your rate, with no extra charge from Orbi.
• Cloud: On Orbi Cloud, "one sentence is enough to get started." The control panel is orbi-build/orbi-cloud (private cloud control plane). Cloud starts from GitHub: log in, install the App, get 3 free trial deliveries, no credit card or subscription required. The cloud runner will clone your authorized repository to its operating machine during delivery.
• Installation and code: The open-source repository is orbi-build/orbi, with an additional website repository orbi-build/orbi-website. The license is AGPL-3.0; if policy does not permit AGPL, the Sustainable Use License can be chosen instead, and self-hosting is free under both licenses.

Use cases

• Teams that want to automatically advance Issues in the backlog to tagged releases.
• Teams that need an independent review step, where reviewers can directly modify the branch rather than just comment.
• Teams that want to retain GitHub's existing branch protection, required checks, and manual approval as the final gate.
• Teams already running Orbi; the official site also lists third-party delivery records merged in other people's repositories.
• Difference from editor-type AI tools: editors make typing faster, while Orbi handles what happens when "no one is typing."

Limitations and notes

• Self-hosting requires preparing Python 3.14, git, gh, systemd, and a model endpoint yourself, and bearing model usage costs (at the provider's rate).
• In cloud mode, the runner will clone your authorized repository to its operating machine during delivery; for code visibility issues, refer to the Cloud page description.
• Specific prices are subject to the official site: 3 free trial merged deliveries; Solo annual billing US$24/month (US$290/year, save 17%; monthly billing US$29), including 400M tokens; Pro annual billing US$66/month (US$790/year, save 17%; monthly billing US$79), including 1.2B tokens; founding partners get 50% off for life, limited to 6 spots.
• The official page shows GitHub 217 stars.
• Information not provided in the text: review pass rate, the specific list of supported models, Windows support, Chinese documentation, SLA , etc., are not yet provided.