Overview
How it works
Commands
Contribute
Install →
MIT licence · 48 agents · 35 commands
public repo opening shortly
Overview
How it works
Commands
Contribute
Install →
MIT licence · 48 agents · 35 commands
public repo opening shortly
One brief in. brew_chai plans the work, asks the few hard questions, splits it into sprints, executes in parallel, audits the result and hands back a PR.
MIT licence · repo opening shortly · 48 agents · 35 commands · What we would take →
~60 MB signed binary · macOS, Linux, Windows · npm and the public repo follow when it opens
A brief goes in. A branch comes back with code, tests, commit history and a description that reads the diff. Work that goes into review, not chat output someone has to assemble.
Plans must score 95/100 with the plan critic before decomposition. Every sprint ends in a tiered audit; the final sprint always gets the full swarm. A task cannot be marked complete with typecheck errors or a failing review.
brew_chai is orchestration. Inference runs through the Claude Code you already have, on your subscription or your API key. Every prompt that touches your repo is a plain file on your machine; read it before you run it.
Phase names map to real commands and agents in the plugin. The rail on the left follows you through them.
~/repo · brew_chai · live
/detect-stack writes .stack-profile.md: language, framework, test runner, lint and build commands every agent reads.
/brew classifies the intent, estimates scope in a few greps, and picks the workflow.
plan-author reads the codebase, asks the clarifying questions that matter, and writes the plan document.
plan-critic scores it 0–100. Under 95 goes back with itemised feedback, up to five rounds.
sprint-planner, task-creator, task-assigner: sprints, task IDs, dependency graph, parallel batches.
task-implementer agents run in parallel batches, in isolated git worktrees, with a checkpoint commit per sprint.
The tiered gate: Tier 1 sanity, Tier 2 parallel auditor swarm, Tier 3 comprehensive. Fix loop until zero issues at threshold.
worktree-merger merges in dependency order; knowledge-extractor proposes rules for next time; docs-syncer updates the docs; you get the branch.
plan-critic passes at 95/100 and not below. The TaskCompleted hook refuses to close a task with typecheck errors or a review under threshold. The threshold is yours to set in .brew_chai.yaml.
/deep-clean audit --tier 2 runs hygiene, completeness, craftsmanship and security at once, and adds performance and design-system auditors when the code in scope calls for them. Each scores 0–100. Tier 3 adds dependencies, environment and knowledge extraction. The final sprint of every feature gets Tier 3.
Fixes run until zero issues at your severity threshold. Circuit breakers: plateau detection, two failed attempts on the same issue triggers investigation, a checkpoint with you every five rounds.
Implementers work in isolated git worktrees on batches the task assigner built from the dependency graph. worktree-merger brings the branches back in order. The orchestrator itself never runs git merge.
run-state.json per feature, state.json per sprint. Re-run the command and it resumes from the last completed phase. A PreCompact hook saves state before context compaction.
Twelve hook scripts on nine Claude Code events. No push, merge or reset --hard during a sprint. Reviewers can only write markdown. Every review must carry a numeric score and a verdict. Files are auto-formatted after every write.
brew_chai is MIT licensed, and the repository opens shortly at github.com/Creative-Human-AI/brew_chai. Every agent, command, hook and rubric is a plain file you can read before you let it near your repo. We built it to ship Glasshouse and this site, and we still use it every day.
We want contributors. The agents are markdown; the hooks are shell; the scripts are Bash and Python that return JSON. If you have shipped production code with coding agents, you already know enough to make one of these better.
The areas below are where help will count most on day one. Tell us which one you would take and we will point you at it when the repo opens. Maintainers review every PR through the same gates the tool runs on itself.
MIT licence · last release v3.0.8 · public repo opening shortly
Copyright 2026 Heiko Altrichter / Creative Human AI.
Free to use, fork and self-host. No paid tier, no seat count, no key to buy. Talk to us →
PreToolUse hook rejects git push, merge, reset --hard and checkouts of main while a sprint is active. Reviewers and critics can write only .md files.Install it in a minute, run /brew on something real, and tell us where it gets in your way.
brew_chai is built by Creative Human AI, an AI lab in Berlin. We use it ourselves, every day, on Glasshouse and on this site.