Ships PRs, not suggestions.
One brief in, one PR out. Code, tests, commit messages, a description that reads like a senior engineer wrote it: actual deliverables that go into review, not chat output someone has to assemble.
Brew your next
feature · bug fix · refactor · migration · audit · integration.
One command, $ /brew_chai:brew "<your intent>": and the autonomous pipeline plans the work, asks the few hard questions, then ships you a PR. Tests included. No mid-run check-ins. Code never leaves your machine.
Three promises that land on the same PR.
Every change ships with tests.
100 / 100 / 100 / 100 coverage on every line brew_chai touches. The anti-defer hook rejects "I'll add tests later", suppressed coverage markers, and TODOs. No exceptions, no waivers without explicit authorisation.
Your code stays on your machine.
brew_chai fetches signed prompt packages, never code. Inference runs through whichever AI IDE your team already uses. Teams plan optionally self-hosts the backend behind your firewall.
Honest fit. Read the right column too.
- You ship production code, not throwaway prototypes
- Code review and tests are non-negotiable on your team
- You want one tool for autonomous work, not a stack of agents to wire together
- You already use Claude Code, Cursor, OpenCode, Continue, Windsurf, or aider
- Your security review will accept signed-prompt-only outbound, or you want it self-hosted entirely
- You want a chat interface to bounce ideas around, that's already what your IDE provides
- You're picking a model and crafting prompts yourself, brew_chai is the layer above that
- You want suggestions and autocomplete, not delivered work
- Your environment cannot accept any outbound traffic, even signed-prompt fetches (Teams self-host fixes this)
One command. One PR. Every time.
$ /brew_chai:brew "add OAuth login (Google + GitHub) to the settings page"
plain English · one command · 47-minute wall time on a medium feature
PR · ready
#1284 · main ← brew/oauth-settings
Add OAuth (Google + GitHub) to /settings
11 files · +427 / −23 · 11 commits · 0 deferrals
- tests, 14 added · 12/12 existing pass · cov 100/100/100/100
- code-reviewer, 96/100 · approved
- security-sentinel, 98/100 · no findings
- deep-clean, 6 auditors, all green
- anti-defer hook: passed (no TODOs, no suppressions)
Eight phases. After Q&A, silence.
Setup
Repo context, conventions, branch posture.
Scope
Classify intent: build, fix, heal, or tool.
Ambiguity
Find what's genuinely uncertain in the brief.
Q&A
One round. Stakes-tagged. Low stakes default-accept.
Plan
Critiqued plan. Reviewer cycles to a numeric gate.
Decompose
Sprints, tasks, dependency graph, agent assignment.
Execute
Parallel agent worktrees · audit-fix loop · coverage gate.
Handoff
PR, summary, escalation log, knowledge extracted.
brew_chai never sees your code.
- Code never leaves the developer's machine.Signed prompt packages flow in. Code stays local. Inference runs through your AI IDE.
- Ed25519-signed prompts.Every prompt fetch is signature-verified. Tampering breaks the chain.
- BYO inference.Claude API · Bedrock · Vertex · Foundry, whatever your team already pays for.
- Air-gap ready (Teams plan).One Docker image holding the prompt backend, deployed inside your firewall.
- Anti-defer commit hook.Rejects suppressed coverage markers, type ignores, lint ignores, and TODOs.
- Dead-drop guarantee.If Creative Human ever shuts down, signed prompts go public so deployed installs keep working.
100/100/100/100 on every changed line.
Every diff brew_chai produces is verified against the new lines it changed. Statements, branches, functions, and lines all hit 100%, or the task does not merge.
Coverage-suppression markers (/* istanbul ignore */, # pragma: no cover) are rejected by the anti-defer commit hook. Bug fixes ship with the regression test that proves the bug existed.
Pre-existing untested code on touched files isn't the run's problem. It feeds the coverage-debt heal backlog for a later /brew_chai:brew "heal --coverage" pass.
// about the name. Creative Human AI sounds like CHAI. CHAI is tea. You brew tea. So we built a thing that brews: not tea, but features. Naming was the fast part. The eight-phase pipeline took longer.