Autonomous coding agent · open source

An open-source autonomous coding agent, from Issue to release

The endpoint is a tagged release, not a merge. Label one GitHub Issue ai-ready; Orbi writes the code, opens the PR, runs an independent review, merges only the reviewed head, then tags and publishes the release — nobody has to be in the room.

Unattended is a property of the checks, not of the model

The mechanics are the ones described in “Claude Code in Actions: who presses merge?” — nothing added.

  1. 01 · TWO SESSIONS
    The reviewer is not the author

    Writing and reviewing are two separate model sessions. The reviewer starts after the pull request exists and reads the frozen diff. It can repair the branch, push, and re-emit its verdict against the new head. The loop is bounded at five rounds, and exhausting them is a stop, not a merge.

  2. 02 · MERGE GATE
    Three conditions re-checked at the moment of merge

    CI is green on the exact reviewed head; the head has not moved since the verdict; the base is current. If any check fails, the delivery stops for a human — it never merges because it looked fine.

  3. 03 · DETERMINISTIC RELEASE
    The release is a state machine, not a model session

    It freezes the base, waits for green CI, bumps the version, tags, and publishes. Any failure stops and waits for a person.

Orbi delivers its own repository

546 pull requests merged, 66 releases tagged — all of it in the open on github.com/orbi-build/orbi. Every review verdict sits in the public Issue threads; both numbers are yours to count.

Write the Issue and the release ticket

The Issue is one runtime outcome, with acceptance criteria something else can check. The release ticket is a version number and a scope. Those two documents are the whole job; between them, the agent runs unattended.

Between those two tickets, nobody is in the room.

Self-host it, or let Cloud run it

Open source.
The repository is github.com/orbi-build/orbi, under AGPL-3.0.

Self-hosted.
A Python runner that works with any OpenAI-compatible API (DeepSeek, for example), or a ChatGPT/Codex subscription login. See the provider docs.

Orbi Cloud.
Sign in with GitHub. The first 3 merged deliveries are free, no card. Solo is US$29/month with 100M included tokens; Pro is US$79/month with 300M. When the quota runs out, new deliveries pause — no overage billing. Each connected repository runs under its own Unix user. Start with Cloud →

What a delivery costs

Across 20 merged pull requests delivered in off-peak hours, the median cost was $0.125 per pull request. The method and the per-PR numbers are on the cost page.

Questions before you hand over the keys

Is it really unattended?
Between the two tickets, yes: implementation, review, merge and release run on their own, and any failed check stops the delivery to wait for you.

Which models does it use?
Any OpenAI-compatible API (DeepSeek, for example), or a ChatGPT/Codex subscription login.

What does it cost?
Self-hosting is free under AGPL-3.0; model usage is billed by your provider. Cloud: first 3 merged deliveries free, no card; Solo US$29/month (100M tokens), Pro US$79/month (300M tokens) — running out pauses new deliveries instead of billing overage. Measured medians: the cost page.

How does it compare to Devin?
We keep a measured, dated side-by-side on the Devin comparison page.

Where does the workflow end?
At a tagged GitHub Release. The Issue-to-release page walks all six steps from the label to the release.

Hand over an Issue, come back to a release

The commits, the review rounds, the merge and the tag are the agent’s work.