dreamteam
AI agents say “done” too easily. dreamteam makes a second agent prove it.
example run · “add OAuth login”
- castweb crew: backend, frontend, security
- producebackend reports 12 tests passing
- gatereality checker re-runs them against a broken stub
- gateneeds work: 3 tests still pass with the code removed
- fixthe tests now check the token, not just the status code
- gatepass: every claim matched to a command and its output
- integratemerged from its own git worktree
-
One loop
Produce, gate, fix, integrate. Building, audits, research and QA all run through it; only the crew changes.
-
An honest report
Every “done” comes with what was checked. Anything it could not check is marked unverified, with what you should check yourself.
-
Six CLIs
Claude Code, Codex, Gemini, CodeWhale, OpenCode and Cursor.
-
Small on purpose
One skill you call per task, not a framework to build on. Open source under Apache 2.0.
Try it read-only
After installing, point it at a repo. The audit profile reproduces every finding against your real code and only prints a report.
/dreamteam --profile audit "find bugs in this repo"