Use cases
Test six hypotheses at once.
Performance work is guess, measure, repeat. Loom runs the loop in parallel: each session profiles one suspect, tries one optimization, and has to prove it before it counts.
The method
Stop optimizing in series.
Parallel hypotheses
The slow query, the chatty endpoint, the oversized bundle, the render storm: brief one suspect per session and the Conductor dispatches all of them at once instead of one per afternoon.
Real terminals, real numbers
Sessions run in native PTYs inside a WebGL terminal, so profilers and benchmark runs are the real thing, streaming live. You can watch any terminal mid-measurement or type into it.
Keep only the wins
Loom verifies work before it counts as done, and you read the diffs afterward. An optimization that cannot show its numbers does not make the cut.
Inside one lane
Every session earns its claim.
Write the brief so that each task ends in a measurement: profile first, change one thing, run the bench again. The mission DAG and live activity strips show which experiment each terminal is on, so six investigations stay legible from one screen. And because auto-accept handles the permission prompts, a long profiling run does not pause waiting for you to press yes.
# hypothesis: /search is doing N+1 queries
$ node --prof server.js
$ autocannon -d 30 http://localhost:3000/search
# change one thing, then run it againClosing the loop
From numbers to merged.
Feel it in the preview
Dev servers are auto-detected and opened in an inline preview tab, so a snappier page is something you click around in, not just a chart. See web preview.
Take changes hunk by hunk
In the editor, AI diffs are accepted or rejected hunk by hunk. Keep the cache, drop the speculative rewrite that came with it.
Pick what works
Some hypotheses lose. Review all six lanes in the git graph, merge the winners, and discard the rest, the same move as prototyping.
Hand it the work.
Walk away.
macOS, Linux, and Windows. Around 13 MB. Free and open source.