You steer.
They build.
Two gates stay yours.
loopd is a Claude Code plugin that runs a small team of AI agents on your repositories. You describe the work. They plan it, build it and get it reviewed. You promote, and you merge.
The same list sits in AWAITING.md, on the session banner and on the board: /pane beside your session, /loopd:board publish on your phone.
Five stops. Two of them are yours.
The agents work in their lane. The flow only drops into yours twice: nothing is dispatched until you promote, and nothing reaches main until you merge.
Shipped to main
Review can send a PR back for a fix: one round per reviewer, then it comes to you.
Not every project is code
A kind=research project produces documents in the bundle: no repo, no PR. Its second gate is you approving the deliverable.
The Crew
From install to your first Pull Request.
Needs Claude Code, git, gh (logged in) and bash. The plugin installs once per machine. Each group of repositories gets its own bundle: a small git repo that holds plans, tasks and notes, never application code.
Two or three people can share one bundle; each project has an owner.
None of them is decoration.
Independent review
Every PR is cleared by a reviewer with fresh context, never the implementing agent’s own report.
Exit 0 is the only clearance
Missing, pending, skipped and unreadable checks all refuse. So does a reviewer that declined to review.
Destructive commands are blocked
A hook stops shapes like terraform destroy, force-pushes to protected branches and rm -rf of a repo root.
Never deletes on a guess
The worktree pruner only reports, and bundle repair is report-only by default. Stale means ask the human.
The gates can be delegated
Only through the separate loopd-yolo plugin, and only for projects set to autonomy: yolo. Uninstall it and every project on this machine is gated again.
How autonomy works →