Turns a labelled GitHub Issue into a reviewed, merged PR and a release
Pick Orbi if your backlog is made of well-specified Issues and you want pull requests that arrive already reviewed, run free on your own machine or on a flat-price cloud plan. Skip it for editor pair programming. Its GitHub repository showed 258 stars on 18 November 2026 and under three months of history, so treat it as early.
An independent review session that can edit the branch, with the merge performed by the Runner against a head-bound verdict.
The review-and-merge loop recorded in the user's own GitHub history is hard to leave once teams rely on it, but the open-source core means a rival can copy the design.
Main trade-off
It is under three months old, so long-run reliability and the maker's staying power are unproven.
Self-hosted core: free under AGPL-3.0 or the Sustainable Use License, you pay your own model provider. Orbi Cloud: first 3 merged pull requests free with no credit card. Solo $29 a month with 400M tokens and 1 repository. Pro $79 a month with 1.2B tokens and 5 repositories. Yearly plans $290 and $790. Founding members get 50% off for life on the first 6 places, which were 5 of 6 left on 18 November 2026. No per-token overage: when the monthly tokens run out, new deliveries pause.
| System | Type |
|---|---|
| GitHub | Native |
| GitHub Actions CI | API |
Orbi is an open-source agent that takes a GitHub Issue labelled ai-ready, writes the code in an isolated worktree, has a second session review and repair the pull request, merges it once CI is green, and publishes tagged releases on request. It is licensed AGPL-3.0 or Sustainable Use, and its repository carries 150 tags.
Issue to merged PR
Adding the ai-ready label makes Orbi claim the Issue, work in an isolated worktree from a frozen base commit, run the test suite and open a pull request that closes the Issue.
Independent review that can repair
A separate session reviews the exact diff, edits the branch, reruns tests and issues a verdict bound to the reviewed head commit; the docs say it can run on a different model.
Merge gate run by the Runner
The Runner merges only with a clean verdict and CI green on the latest base, uses a match-head-commit check, and the AI itself never pushes protected branches.
Release tasks
A release Issue runs the gates, waits for CI on the release commit, tags the version and publishes notes listing every Issue in scope, as in the v0.7.48 release.
Steering and stop questions
A correction comment from a user with write access restarts the run with it, and a delivery that needs a decision stops with a structured question and the ai-blocked label.
GitHub as the only ledger
Issues, labels, comments and PRs carry the full record with no database, and every log line and progress comment shares one run_id.
| Agent | Where it runs | Price | Maturity | Rating |
|---|---|---|---|---|
| Cloud · Your machine | Free plan | Beta | 3.4 | |
| Cloud · Your machine · API | Free plan | Available | -- | |
| Cloud · API | Free plan | Available | -- | |
| Cloud · Your machine · API | Free plan | Enterprise | -- |
Orbi Cloud has two plans on its pricing page, read on 18 November 2026. Solo is US$29 a month for 400M tokens and 1 repository, and Pro is US$79 a month for 1.2B tokens and 5 repositories, with yearly billing at US$290 and US$790. The first 6 founding members pay half for life, US$14.50 and US$39.50, and 5 of 6 places were left. There is no per-token overage, so deliveries pause when the allowance runs out.
Yes, two ways. Orbi Cloud merges your first 3 pull requests free with no credit card, and the self-hosted core is free forever for personal or internal company use under either licence. A self-hosted run still bills you at your own model provider's rate, since Orbi adds no fee on top.
Orbi's Runner merges a pull request itself, but only after a clean review verdict tied to the reviewed commit and with CI green on the latest base. The AI never pushes protected branches, and your branch protection and required approvals still apply. A delivery that cannot be resolved stops with the ai-blocked label for a person, and a correction comment from someone with write access restarts it.
Self-hosting lets you bring your own model endpoint, and the review session can be pointed at a second model. Cloud defaults to DeepSeek's deepseek-flash unless you add your own key, and that provider receives the prompts and repository context. Hosted workers run at Vultr and delete a finished working copy after 120 quiet minutes.
You need a machine that stays on, running Linux with systemd or macOS with launchd, plus git, the GitHub CLI version 2.94 or newer, curl, uv and Python 3.14 or newer, which uv can provide. The coding agent needs Node 22.19 or newer, and no GPU is required. Its maker lists macOS as unverified on real hardware.
Orbi is an open-source agent that works through a GitHub backlog on its own. You write the goal and acceptance criteria in an Issue and add the label ai-ready; Orbi claims it, writes the code in an isolated worktree, runs the project's real test suite, opens a pull request, and takes that PR through independent review, repair and merge. It is an agent and not an editor tool because nobody types while it works: the only human inputs are the Issue, an optional correction comment, and the request for a release. Its public repository started on 24 August 2026 and was published under AGPL-3.0, with the Sustainable Use License as an alternative; the site footer names Low Code, LLC as its publisher, and Hokai has no record for that company yet.
A second session that did not write the code reviews the exact diff, can change the branch and rerun the tests, and judges the repaired head. The Runner, not the model, performs the merge, and only when CI is green on the latest base. Self-hosted Orbi drives a coding agent called Pi through any OpenAI-compatible API or a local model, and the reviewer can run on a different model from the implementer. Orbi Cloud runs the same loop on hosted workers and, unless you add your own key, defaults to DeepSeek's deepseek-flash model. All state sits in GitHub Issues, labels, comments and PRs, so a crashed run resumes from the same branch and pull request.
It fits teams and solo maintainers whose backlog is made of well-specified tickets: bug fixes, documentation, small features, dependency work, and chores that a reviewer can check against written acceptance criteria. Third-party pull requests listed on its site were merged in other people's repositories, which is the closest thing to an outside signal the project publishes so far, and the claims there are the maker's own.
On 18 November 2026 Hokai read the v0.7.48 release page on GitHub: it lists two closed Issues in scope, five CI checks that all passed after a 390-second wait, and the release commit f405e1d, published by the orbi-build account. Releases are frequent, with v0.7.46 and v0.7.47 published the day before, mostly reliability fixes to how a stopped delivery asks a person a question and carries out the answer.
Skip Orbi if you want a pair programmer inside an editor, or if your tickets are vague. Claude Code suits interactive work at the keyboard, and OpenHands offers a similar open-source background agent with a larger user base. If you want an agent managed by a large vendor, Jules from Google and OpenAI Codex also take a task and return a pull request, and Devin is the commercial agent Orbi's own repository names as the product it is an alternative to. The wider list is in the coding agents directory. Orbi works only on GitHub repositories. Choose Orbi when you have a written backlog and a CI suite you trust; otherwise pick one of the others.
Autonomy, guardrails, integrations and pricing on this page are read from Orbi's own documentation. The verdict is Hokai's reading of them, not the vendor's claim.