Agent cockpit for spec-and-loop driven development.
Start with a repository, make a plan, run your agents, and review their work.
Write a PRD, break it into tasks, and let coding agents loop on them until every task is accepted. 🚀
Loopfleet is a native macOS app for scheduled and continuous task execution. Choose Claude, Codex, pi, or Cursor for each run. Select a model and a schedule. The agents work through your task list until you accept each task.
Context is kept deliberately small per run, so each agent stays focused on its current task instead of drowning in accumulated history. The app keeps the plan, the runs, and the resulting diffs in one place: scan the fleet, review what changed, and accept or reject tasks without losing the thread.
Agents verify implementation through executable tests and repository checks. These can include component, interface, browser, end-to-end, and system tests.
Loopfleet does not provide computer-use agents or desktop-control automation. Test harnesses can launch the application under test. Manual clicking and visual inspection belong to human release review.
Generated PRDs keep manual release review outside executable tasks. For existing PRDs, agents record manual checks as pending human review. They must pass required automated checks before task completion. They must not report manual checks as passed.
Write a normal PRD with tasks and acceptance criteria. Loopfleet supplies the full plan reference and runner instructions automatically.
Agents change files. Loopfleet owns Git snapshots and commits, reruns verification, and marks successful work Ready for review. Missing external prerequisites produce a resumable Blocked run rather than repeated failed attempts.
Configure mandatory preflight and verification commands once per project in .loopfleet.json. See the execution guide for setup, Docker prerequisites, evidence, and recovery.
While a run is active, click Tap in the run view or press ⌘⇧T (Ctrl+Shift+T) to open the composer and send guidance. Capable agents can receive live steering during a pass; other agents queue notes for the next pass. Notes do not start an extra pass if the run ends. Check the run's tap status for delivery: sending alone does not confirm delivery. Guidance can also be queued when live steering is unavailable, or delivery can be reported as unknown.
Live steering requires Codex 0.154.0 or later (app-server) or
pi 0.80.3 or later (RPC). Older, unrecognized, or unavailable versions
use notes instead of live steering. Discovery and run handles use the same
version check. Claude Code and Cursor keep their existing transports and
use notes (can_steer() == false).
Download Loopfleet 0.3.17 — Apple Silicon · Intel
Signed and notarized .dmg builds; the app updates itself from there.
Older builds: all releases.
Or install with Homebrew:
brew install --cask matija/tap/loopfleetWIP.
See build/README.md.