Use cases
One refactor, six sessions.
Cross-cutting refactors stall when a single session has to hold the whole change. Loom splits the work into a task DAG, gives each seam to its own Claude Code session, and verifies every piece before it counts.
How it splits
From one giant diff to six bounded ones.
Plan first
The Conductor turns your goal into a mission DAG, so shared types land before their consumers and renames land before call sites. Ordered where it must be, parallel everywhere else.
One seam per session
Each of six real Claude Code CLI sessions takes a bounded slice of the codebase. Live activity strips show what every terminal is doing right now, and you can type into any of them to redirect.
Verified before done
Loom re-checks each task before it counts as finished. A session that declares victory early gets caught, not merged.
The brief
Describe the end state, not the steps.
You write one goal. The Conductor decides the split, dispatches tasks across the fleet, answers permission prompts, and nudges any session that drifts. If a CLI exits or hits a rate limit, Loom recovers it and re-sends the task.
# the goal you type
$ Replace the legacy EventBus with typed channels
# roughly how the Conductor plans it
# 1. define channel types in core/
# 2-5. migrate api/, web/, workers/, cli/
# 6. delete EventBus, sweep for stragglersQuestions
Refactoring with a fleet.
What if two sessions need the same file?
The Conductor plans tasks along separate seams and sequences dependent work in the DAG. Because every session is a real terminal, you can also jump in and redirect one mid-flight. The final say is yours: the source control panel shows the full git graph of what the fleet changed before anything ships.
Do I review one giant diff at the end?
No. Changes land per task, so you review them as bounded diffs, and the editor lets you accept or reject AI edits hunk by hunk. See reviewing fleet work for the full flow.
Hand it the work.
Walk away.
macOS, Linux, and Windows. Around 13 MB. Free and open source.