wtf is brigade?

Brigade is a multi-agent build orchestrator packaged as a skill for Claude Code, built around the idea of splitting one big coding task across several AI agents that each play a different role, the way a professional kitchen splits cooking across different stations instead of asking one person to do everything. In Brigade's own terms, a strong model called the Chef reads the task and writes a set of tickets, which are self-contained sub-tasks, but never writes the code itself. Those tickets are handed to a group of cheaper or faster models called cooks, which can run on Anthropic, OpenAI-compatible, or Ollama models, and each cook works in its own git worktree when the work needs to be isolated from the others. Once a cook finishes, a separate reviewer model called the Pass looks only at the actual code diff and checks it against the acceptance criteria, without ever seeing the cook's explanation of what it did. If the Pass rejects the work, it goes back to the same cook, with its previous context kept, for up to two more attempts before it gets escalated to the user. A final step called Service runs an end to end check: building, testing, and integrating everything together. The stated reasoning is that big tasks tend to fail in a single agent because the context window fills up with implementation details, and because the same model that writes the code also ends up grading its own work. Splitting planning, cooking, and reviewing into separate contexts is meant to fix both problems, and using cheaper models for the repetitive cooking work while reserving the strongest model for planning is meant to save money too. Brigade has a companion project called Cerebro, a separate skill that writes reference documentation for a repository, and the two can work together automatically if both are installed, though neither requires the other. Installation is a git clone into a Claude Code skills folder, or running an included install script, after which the skill is triggered with a slash command or natural phrasing inside Claude Code. Configuration, including which model plays which role, is stored in a per-repo JSON config file. The project requires Python 3.9 or newer and is released under the MIT license.