Metaminds ResearchAI Dark Factory · Open source
uzi: an AI dark factory.
Specs in. Pull requests out. An open-source factory for software that runs with the lights off — label an issue, approve the plan, and it opens a reviewed pull request, never touching main.
completed
spent
licence
01The idea
Issues in,
pull requests out.
Even with today’s agents, “AI coding” keeps you in the driver’s seat of a single session — you kick it off, watch it work, nudge it when it drifts, and start the next task yourself. You are the conveyor belt.
uzi inverts that. The unit of work is an issue, not a message. It reads the issue, plans the change, builds it, reviews it, and opens a pull request from a new branch — while you stay out of the loop until the two decisions that matter.
02The pipeline
Two decisions stay human.
Everything between is the floor.
03The crew
A lead that delegates, and specialists that check its work
Work is not done just because an agent says so. The lead orchestrates; a role-based crew implements and validates in parallel, looping until the change holds up.
Plans the run and delegates instead of doing everything itself, working the approved plan one milestone at a time.
Writes the change for the current milestone, committing each as its own reviewed slice.
Read the diff for correctness and for security, reuse and simplification, fanning out in parallel.
Exercise the change and check its claims against the codebase, looping back to the coder until it holds.
04Accountable
Every run is reviewed, measured and costed
Nothing about a run is a black box. It is scored after the fact, tracked while it runs, and accounted for to the token.
05Where it stands
Alpha, and it works —
building its own next version.
And the fun part: uzi builds uzi. Issues get filed, plans approved and PRs opened — so a growing share of the factory is written by itself while a human reviews.
06Scheduled jobs
The factory works the factory
Pointed at its own repo, uzi’s standing automations add up to a self-improvement loop — hunting its own bugs, strengthening its tests, keeping its docs honest, and proposing its next feature. Every job falls back to a plain report when it has nothing worth landing, so a quiet week produces no empty pull requests.
| Schedule | What it does | Cadence |
|---|---|---|
| bug-triage | sweeps bug-labeled issues | daily |
| planned-sweep | sweeps Planned-labeled issues | daily |
| docs-hygiene | mechanical documentation fixes | weekly |
| test-improvement | lands new tests only, no production code | weekly |
| bug-hunt | a deep audit of one subsystem, one focused fix | weekly |
| self-improve | scans the codebase and opens a self-improvement PR | ~2 days |
| feature-bingo | brainstorms one new feature and proposes it | weekly |
feature-bingo is the factory designing its own next machine.
Once a week it reads the existing ideas, checks what already exists so it doesn’t repeat itself, and proposes exactly one concrete new feature — the problem it solves, a sketch of how it works, and where it lives — as a pull request. A chunk of uzi’s roadmap arrives as PRs to wake up to.
07Watch it anywhere
Follow the floor from
the terminal or your phone.
uzi tui — the crew, milestones, rate-limit meters and the live transcript, in your terminal.
08Where it came from
An AI research project, released as open source
uzi began as an AI research initiative at Metaminds and has been built on both work and personal time since. Metaminds takes open source seriously, so releasing it was an easy call. It is MIT licensed — a helm install away on Kubernetes, or a docker compose up on a laptop.
Use at your own risk.
uzi runs autonomous agents that read your code, run commands inside their workers, and open pull requests using your own model tokens. Run it against repositories you own. You stay in control by design: review the plan before you approve it and the diff before you merge it — uzi opens pull requests but never merges them and never touches main.