A dark factory that runs on your Mac: GitHub issues in, shipped work out, across several repos, with almost no human in the loop.
Every 10 minutes a small Node script looks at your GitHub issues, your PRs and your running Claude Code sessions, then starts or wakes a Claude session to move each issue one step forward. The script itself makes no model calls, and GitHub is the only state of record: factory: * labels plus one machine-owned comment per claimed issue.
untriaged ──triage──► factory: ready ──build──► factory: working ──PR──► factory: in review ──you merge──► closed
│
└── stuck? one question on the issue, you reply, it resumes
A repo can also skip the PR stage and have workers land reviewed commits straight on its base branch, or collect a whole project's issues on one long-lived branch that ends in a single PR.
| Piece | What it does |
|---|---|
| Claude Code | every worker is a background claude --bg session in its own git worktree |
GitHub issues + gh |
the backlog, the factory's state (labels and one state comment per issue), PRs, CI, sub-issues and blocked-by relations |
macOS launchd |
runs the loop every 10 minutes, keeps the dashboard up at http://localhost:7788, runs the monitor every 30 minutes |
| UI Verify | optional visual check: the agent sees what its UI change did and fixes it before a human looks |
All in skills/, linked into ~/.claude/skills/ by setup.sh.
The loop (the same skills as UI Verify's uiverify-workflow plugin). A repo can carry its own versions in .claude/skills/; workers use the repo's when it has them.
| Skill | What it does |
|---|---|
factory |
one change from ask to shipped: design if needed, implement, review, verify, then a PR or a push |
design |
drafts a design and has two models critique it until it holds up |
review-loop, review-multi-model |
Claude and Codex review the diff, verified fixes get applied, repeat until clean |
e2e-verify |
drives the change through the real app with Playwright |
babysit-pr |
watches CI and review comments on a PR and fixes what fails |
add-rule |
turns a correction into a CLAUDE.md rule plus a lint guard, so it never happens again |
evaluate-run |
grades a finished run against your conventions |
The factory (what runs it unattended):
| Skill | What it does |
|---|---|
factory-auto |
factory with no questions: decides, records each choice under "Decisions taken", parks on the issue only for the stop reasons |
factory-triage |
reads a new issue: ready, split into a batch, needs a decision, needs a human, or later |
factory-plan |
groups ready issues into work packages, one worker each |
factory-project |
the one place you spend real time: spec, low-fidelity flows and wireframes, architecture and the issue batch, together |
factory-run |
start, stop, pause, status, add a repo or a target, in plain words |
factory-monitor |
read-only health check every 30 minutes, tells you when the loop quietly stalls |
factory-ops |
an operator session: loads how the factory works, checks its health, helps you change it |
For visual testing, add UI Verify's own skills: npx skills add uiverify/uiverify.
git clone https://github.com/uiverify/factory ~/.claude/factory
cd ~/.claude/factory && sh bin/setup.shThen edit factory.json: your name and GitHub login under owner, each repo (GitHub slug, local clone, base branch, ship mode) and one target. Rerun setup.sh to get a principles/<repo>.md per repo and fill it in; principles/common.md holds the rules every repo shares. Those files are the important part: they are the judgment workers use instead of asking you (what to decide alone, when to stop, how the repo ships). Finally, add @~/.claude/factory/issues.md to each repo's CLAUDE.md so every agent there, factory or not, follows the same issue flow.
Try one tick without side effects, then start it:
node bin/factory-tick.mjs --dry-run --verbose
sh bin/install-launchd.sh 600 # or /factory-run start in Claude CodeStart small: one target, one or two issues in flight, and every change going through a PR you merge.
- This is one person's working setup, published as is. Not a polished tool. Expect to read the code and adapt it.
- Workers run with no permission prompts (
--dangerously-skip-permissions, andbypassPermissionsin each worktree). A worker can run any command as you, with your credentials. Run it only on a machine and accounts where that is fine. - macOS, Claude Code and GitHub only.
- It is only as good as your repo. Conventions in CLAUDE.md, lint and tests that catch mistakes, a way to see UI changes. The factory runs the loop; your repo supplies the judgment.
- It uses your Claude subscription. Several Opus sessions at once is heavy. The dashboard's Usage page shows the live weekly allowance and what each issue spent; per repo or per target you can run workers on Sonnet instead.
How it works in detail, every config key, and the known quirks: docs/how-it-works.md.
MIT licensed.