Instrument  /  Open orchestration engine  /  Module 02

The engine
is open.

Brew Chai is the open-source orchestration engine we use to build Glasshouse — and everything else Creative Human ships. Apache 2.0. Maintained by Creative Human engineering. Read the code that built the rehearsal you're about to evaluate.

01  /  Why it's open

Open-sourcing Brew Chai is not a values gesture.

Glasshouse rehearses consequential decisions for Fortune 500 CEOs, Chief Communications Officers, ministry directors, and institutional fund managers. Every one of them will defend the rehearsal to their own board, to their compliance team, to their regulator, or to their own conscience. The defense has to survive a specific question: "how do you know the rehearsal is trustworthy?"

Handwaving about "our proprietary engine" does not survive that question. Neither does "we tested it internally." The only answer that survives a board-grade challenge is one the board can verify independently. Which means every layer of the product has to be inspectable.

That is why Glasshouse's trust commitment runs at four levels — claim, run, persona, engine. The first three are software features. The fourth is a licensing decision. All four are required for the no black boxes promise to be structurally true. Stopping the ladder at three would leave a hole at exactly the level the board question lands.

Brew Chai is open because "you can't ask why of a closed engine." We opened it because the product promise requires it. The community benefit is a consequence.

02  /  The /brew flow

What Brew Chai does.

Brew Chai takes a natural-language feature request and routes it through a structured orchestration pattern: intent analysis → workflow selection → parallel agent activation → quality gates → knowledge synchronization. A concrete example:

terminal / brew-chai
# Natural-language entrypoint $ /brew add OAuth2 authentication # Brew Chai plans → implements → reviews → audits → verifies → syncs intent-analyzer (routing to /plan-this) plan-author (draft plan) plan-critic (score: 94/100) [quality gate] task-creator (12 tasks, 4 dependencies) task-implementer (executing...) code-reviewer (PR ready) [quality gate] security-auditor (OWASP Top 10 clear) [quality gate] cleaner (lint + types pass) knowledge-assembler (rules synced) # Diff ready to merge.

Brew Chai plans the change, implements it, runs code review, runs security audit, runs type and lint checks, runs the test suite, verifies the accessibility and performance surface hasn't regressed, synchronizes the knowledge base with what was learned, and produces a diff ready to merge — with explicit quality gates between every phase and a structured trace of every agent activation along the way.

The differentiator is not "it calls agents." Plenty of systems call agents. The differentiator is "it passes work through quality gates before anything lands." The orchestration-versus-automation distinction lives in the gates. An automation pipeline fires agents in sequence and hopes the output is good. An orchestration engine runs agents and then checks their work before moving to the next phase. Brew Chai is the second thing.

03  /  Agent library

48 specialized agents.

Grouped by category. Each category has a handful of named exemplars on this page. The full list, with per-agent documentation, lives in the GitHub repository.

Category / 01

Planning

  • plan-author — transforms a raw feature idea into a comprehensive plan document ready for critique
  • plan-critic — quality gate that scores plans 0–100 and returns them for revision if they don't pass
  • sprint-planner — breaks approved plans into experiment-driven sprints

Category / 02

Engineering

  • task-implementer — systematically executes task batches from a tracker
  • task-creator — converts plans into structured, actionable task specifications with dependency mapping
  • cleaner — runs lint, typecheck, and build fixes

Category / 03

Quality

  • sprint-verifier — lightweight diff-scoped intent checker that catches logic errors and architectural mismatches after major steps
  • craftsmanship-auditor — scans for code duplication, unnecessary complexity, and elegance issues
  • completeness-auditor — scans for missing implementations, stubs, TODOs, and coverage gaps
  • security-auditor — OWASP Top 10, secret scanning, auth flow analysis, injection vectors

Category / 04

Design

  • design-system-auditor — scans frontend code for design system violations
  • accessibility-auditor — WCAG 2.1 AA audit via browser automation
  • frontend-performance-auditor — bundle bloat, re-renders, code splitting, image optimization

Category / 05

Data and GenAI

  • genai-process-mapper — discovers and maps a GenAI process end-to-end
  • genai-process-tuner — proposes and implements targeted improvements to GenAI processes
  • data-flow-tracer — traces how a data value gets created end-to-end

Category / 06

Knowledge

  • knowledge-extractor — classifies sprint insights as rules, guides, or history
  • knowledge-assembler — deduplicates and merges knowledge across rules and guides
  • knowledge-healer — migrates legacy knowledge into the three-layer system

04  /  Workflow library

35 named workflows.

Each one orchestrates a group of agents behind a single natural-language entrypoint.

Full workflows

  • plan-this — the full planning pipeline with a critique loop
  • run-sprints — end-to-end sprint execution with tiered audit gates
  • fix-this — investigate and fix a bug from root cause through implementation and review
  • fix-output — outcome-driven data pipeline fixer that iterates until the output is correct
  • fix-ui-loop — end-to-end UI testing pipeline with detection, execution, analysis, and fixes
  • tune-genai — iterative GenAI process tuning with baseline, evaluation, and refinement cycles

Auditing workflows

  • deep-clean — tiered audits with audit-fix loops and 5-auditor swarms
  • audit-a11y — WCAG 2.1 AA audit on a live URL
  • audit-performance — Core Web Vitals audit on a live URL
  • audit-seo — SEO technical audit on a live URL

Development workflows

  • dev-skill — create, update, or refactor a skill with convention enforcement
  • dev-agent — create, update, or refactor an agent with dependency validation

Knowledge workflows

  • sync-knowledge — deduplicate, merge, and split knowledge across rules and guides
  • verify-knowledge — verify rules and guides against the actual codebase
  • heal-knowledge — migrate legacy sprint knowledge into the current system

Smart router

  • brew — analyzes your natural-language request and routes to the right workflow

05  /  Architecture

Three layers.

Brew Chai is the orchestration substrate. Underneath the substrate, the actual work happens via a runtime that calls the 48 agents with per-agent model assignments and passes their outputs through a quality-gate graph. The stack has three layers.

Layer 01

The router

/brew <natural-language request> enters the router. The router analyzes intent, selects the right workflow, and fans out into parallel agent activations.

Layer 02

The agent graph

Each agent operates inside its own context window with a focused prompt, a specific model assignment, and a narrow set of tools. Agents produce typed outputs that flow forward in the graph.

Layer 03

The quality gates

Between every phase, a quality gate runs — a code review, a security audit, a type check, an accessibility pass, a coverage check. Gates fail loud. A failing gate halts the workflow and returns control to the operator with a structured explanation of what failed and why.

The operator (the human at the keyboard) sees the trace of every agent activation, every gate result, every decision the router made, and every piece of output that flowed through the graph. The engine is not a black box for the operator either — it is a recorded, inspectable process from natural-language request to merged diff.

06  /  Dogfood

How Creative Human uses Brew Chai.

Brew Chai is the engine we ship our own work with. Glasshouse was built using Brew Chai's plan → sprint → audit workflow pattern with explicit quality gates at every phase. Our engineering team runs Brew Chai daily against its own code.

Every plan document, sprint tracker, audit report, code review, and quality gate in the Glasshouse engineering history was produced by Brew Chai running against our own engineering work. The "we eat our own dog food" claim is made publicly, in code, in the open repository — readable for anyone evaluating Glasshouse before they commit to an engagement. You can see the engine at work in our own delivery pipeline before you ever talk to us.

This is not a claim other AI systems companies can make while keeping their orchestration stack proprietary. We can make it because we opened the stack.

07  /  Governance

How the project is maintained.

Maintained by Creative Human engineering. Public roadmap on GitHub. External contributions welcome under DCO — the Developer Certificate of Origin, a single-line affirmation in every commit that the contribution is the contributor's own work to give. No CLA. No corporate signing ceremony. No surprise proprietary relicensing.

Apache 2.0 governs the code. The Brew Chai name is a trademark of Creative Human and is governed by a trademark policy that restricts name usage in ways designed to prevent downstream projects from being rerouted into the applications the Creative Human ethical redlines forbid — electoral politics, military, defense, intelligence, dual-use rerouting. See the trademark policy in the repository.

08  /  For enterprise teams

Need help deploying Brew Chai inside your organization?

We scope bespoke Brew Chai deployment engagements case-by-case. Security review, per-team customization, integration with your existing engineering infrastructure, Advisory around agent-governance decisions. If you want to run Brew Chai at enterprise scale inside your organization and want Creative Human to help land it, start a conversation.