ApolloOffice

A small office of AI workers that turns GitHub issues into verified pull requests, on your own computer and with your own accounts.

Website

Add any GitHub repository. ApolloOffice picks up its bug reports (or hunts for bugs itself), reproduces them, writes a fix with a regression test, proves the test fails without the fix and passes with it, runs the project's tests, reviews the change and opens a pull request. Nothing goes through a server of ours.

Warning

ApolloOffice is in beta. Expect bugs, and please report them.

Note

Every pull request waits for your approval by default, with the diff, the proof and an editable PR text. You can switch a repository to submit on its own.

For optional free-model authors with Codex/Claude subscription review, see the free-first guide.

What's inside

• A fixed pipeline of small workers: Scout, Guard, Vet, Hunter, Repro, Fixer, CI, Reviewer, Submit and Watcher, each at its own desk.
• Proof, not trust. ApolloOffice itself checks that the test is red without the fix and green with it, and that your test command passes.
• Any provider per worker. Claude Code or Codex with your own login, or Claude, OpenAI and Gemini with an API key, or a local model through Ollama. Click the gear on a worker's card to change it.
• Docker isolation on this computer or any machine you reach with ssh. The GitHub token never leaves your computer.
• Watcher follows your open PRs: merges, review requests from maintainers and merge conflicts, which it rebases on its own.
• API for bug reports from your own tools (error trackers, CI, scripts): POST /v1/report with a token from Settings. Reports go through Guard and Vet like issues.
• Error trackers and CI: Sentry and GlitchTip alerts arrive through a webhook link from Settings, and a failed GitHub Actions run can report itself (see below). The same error lands in one report.
• Costs per job, worker and repository, a daily budget in USD for API keys (overall or per repository), and a table of which model gets merged how often at what cost, with a hint when a cheaper one does as well.
• Sign in with a one-time link or your Google account, notifications by ntfy or Discord, and automatic updates.

Install

You need git (on Windows: Git for Windows), the GitHub CLI logged in with gh auth login, and at least one provider: Claude Code or Codex logged in, an API key, or Ollama. Docker is recommended.

Download the binary for your system from the latest release and run it:

apollo-office

It prints a one-time login link. Open it and follow the setup: add a machine, add a repository, check its test command.

apollo-office login-link      # new login link (10 minutes)
apollo-office autostart on    # start at login (Windows, macOS, Linux)
apollo-office status

Options: --port (default 47300), --data <folder>.

Report failed CI runs

Add a workflow that sends failed runs of your CI on the default branch to the office. The office has to be reachable from GitHub, for example through a tunnel. Store its address and an API token as repository secrets.

on:
  workflow_run:
    workflows: [CI]
    types: [completed]
jobs:
  report:
    if: github.event.workflow_run.conclusion == 'failure'
    runs-on: ubuntu-latest
    permissions:
      actions: read
    steps:
      - uses: Silvertree2010/ApolloOffice/report@main
        with:
          url: ${{ secrets.APOLLO_OFFICE_URL }}
          token: ${{ secrets.APOLLO_OFFICE_TOKEN }}

AI disclosure

Every commit carries Co-Authored-By trailers for the models that worked on it, and every PR text says it was generated with AI by ApolloOffice. This is always on: many projects require it, and hiding it can get you banned.

License

AGPL-3.0