Use cases
Make CI boring again.
Flaky tests, slow builds, and broken pipelines are perfect fleet work: parallel, repetitive, and verifiable. Loom drives six real Claude Code sessions through the whole list while you do something better.
The work
Three jobs nobody volunteers for.
Flaky tests
Each session takes one intermittent failure, reproduces it under repetition, and fixes the root cause instead of wrapping it in retries. The fix is re-checked before the task counts as done.
Slow builds
Six hypotheses at once: caching, dependency pruning, splitting steps. Each session tries one change and measures it, and you keep whatever actually moves the number.
Broken pipelines
A session reads the failing logs, patches the workflow, and runs it again. The Conductor watches the terminal and re-dispatches the task if the session stalls or exits.
One brief
Type the goal, get a mission.
The Conductor, a reasoning model you bring your own key for, plans the goal into a DAG of tasks and dispatches them across the fleet. Auto-accept presses the safe affirmative on permission prompts so six sessions never sit waiting, and live activity strips show which failure each terminal is chasing. CI work also queues well after hours, so pair this with overnight runs.
# one brief, six sessions
Find every test in this repo that fails
intermittently. Reproduce each one under
repetition and fix the root cause, no
retry wrappers. Then profile CI and cut
the three slowest steps.Questions
The fine print.
What happens when a test run hangs?
Smart steering watches every terminal. It nudges drifting sessions, relaunches exited CLIs, and re-sends the task after recovery, so one hung run does not end the mission.
Does one passing run count as fixed?
Not by itself. Work is re-checked before it counts as done, and a premature done gets caught and sent back. See verification for how that works.
Do I need an API key for the fleet?
No. The fleet runs on your existing Claude login through the official CLI. Only the Conductor is bring-your-own-key, and it can even be a local model. Details on the Claude login page.
Hand it the work.
Walk away.
macOS, Linux, and Windows. Around 13 MB. Free and open source.