DiffSwarm spawns a fleet of AI agents against your diff. Each agent hunts independently for different bug classes, then findings are cross-verified and merged into a single structured report. Runs locally with your own Codex or Claude Code credentials.
$ brew tap bro4all/diffswarm && brew install diffswarm
$ diffswarm login
Email: [email protected]
One-time code sent. Enter code: ******
Logged in.
$ diffswarm pr https://github.com/acme/api/pull/847 --engine codex --out report.md
Setting up hunt in ~/.cache/diffswarm/runs/20260210-1423-a8f3c1
Loaded diff: 14 files, 847 additions, 203 deletions
Spawning 10 bug_finder agents...
[18s] Bugs: 2 (pending=0 verifying=1 confirmed=1 refuted=3)
[34s] Bugs: 3 (pending=0 verifying=0 confirmed=3 refuted=5)
Hunt finished. Report written to report.mdMultiple agents analyze your diff in parallel. Each one specializes in a different class of defect — concurrency, edge cases, security, logic errors.
Candidate bugs are verified by a separate judge agent. Only confirmed findings with line-level evidence make it into the final report.
Your code stays on your machine. DiffSwarm orchestrates local Codex or Claude Code processes — no proxy, no hosted inference.
Pass --engine codex or --engine claude. Use whichever AI CLI you already have credentials for.
Three profiles: cheap, balanced, thorough. Controls fleet size, duration cap, and agent reasoning effort.
Use --comment to post findings to the PR. Use --dry-run for pipeline smoke tests without LLM calls.