MUSTER

OPERATIONAL

  • open source
  • runs in Claude Code
  • v4
  • MIT

Ship a product with an AI team.

PM
  • Developer
  • UI/UX
  • QA
  • Content
  • Marketing
  • Legal
  • Research

8 AI agents · 1 operator

THIS SITE · SPEC → LIVE VERIFY

curl -fsSL https://raw.githubusercontent.com/thinkArhant/muster-ai/main/scripts/setup-project.sh | bash -s my-product

§02 · WATCH IT SHIP

BODH · SPRINT 4 — CONDENSED FROM THE REAL BUILD LOG

RUN

  1. 20:38 muster sprint/auto · queue read · 8 roles standing by
  2. 20:38 ui-ux production spec + assets · HO-027 filed
  3. 20:46 dev web/ built · rotation port parity ✓ (UTC+ / UTC−) · HO-028
  4. 20:59 pm ✓ independent re-verify · own screenshots · match
  5. 21:06 content copy on-voice · one meta edit · HO-029
  6. 21:11 legal data-not-collected · confirmed at code level · HO-030
  7. 21:16 mkt share card written · HO-031
  8. 21:20 qa re-deriving rotation · 24 dates × 3 timezones …
  9. 21:35 qa 11/11 acceptance · PASS · zero bugs · HO-032
  10. 21:35 pm handoffs accepted · deploy packet → founder
  11. 21:43 gate Role: halt · awaiting operator
  12. ───── deploy bodh.day · LIVE
  1. A real run, replayed: Bodh's website wave — one evening, AI agents, unattended. The queue advances; UI/UX files the production spec.
  2. The developer builds the deployable site — an exact port of the app's date-rotation logic, verified on both sides of UTC.
  3. One agent auditing another: the PM re-checks the developer's work with its own screenshots — never trusting the developer's. Accepted.
  4. Content polishes the copy — on-voice, one meta-description edit.
  5. Legal confirms 'Data Not Collected' in code, not in a policy.
  6. Marketing writes the share card — 134 characters, no competitor mention.
  7. QA doesn't re-run the developer's tests — it re-derives the date math with its own formula, across 24 dates and 3 timezones.
  8. PASS — 11/11, zero bugs. The PM accepts, the same instant.
  9. The operator planned the sprint, left the agents running, and returns to a deploy-ready site.
  10. bodh.day, live. The deploy landed three days later — the gate waited on Apple. This wave: 64 minutes of agent work across 8 sessions, 289 API calls, $24.73 at API list price.

~64 MIN AGENT WORK · 289 API CALLS · $24.73 BODH SPRINT 4 · WEBSITE WAVE ONLY

§03 · THE INSIGHT

The product grows. The briefing doesn't.

Muster puts eight AI roles under one human operator, coordinating through markdown files and a queue. Used bare, Claude Code makes you the memory: every session starts with re-explaining, and the re-explaining grows with the product. Muster moves that job into the system. Decisions land in a knowledge base; each session reads a curated slice — a small always-on tier, the rest on demand, most never — context engineering, per role. The product grows; what a session reads stays bounded. Mechanics run as scripts, judgment in prose — it holds on cheap models.

§04 · THE DECISIONS

  1. I optimized what each agent reads, not how they talk.

    framework — 2026-04-24
    Decision
    Each agent gets a curated brief: what its task needs, nothing more.
    Problem
    What breaks isn't the talking — it's each agent reading too much.
    Trade-off
    Questions between roles travel as files — written, routed, auditable — and each costs a session.
    Mechanism
    Three reading tiers; the PM routes context.

  2. Any rule a script can check, a script enforces.

    framework — 2026-06-13
    Decision
    Mechanics live in scripts; judgment stays in prose.
    Problem
    AI follows written instructions unevenly — one model skipped a required step.
    Trade-off
    Hard floors can't flex: a legitimate exception still trips the gate.
    Mechanism
    Automated gates hold the floors, so premium models buy judgment, not correctness.

  3. I capped every file the agents read — before almost anything else.

    framework — 2026-04-12, first commit
    Decision
    Every file the agents read has a hard size cap.
    Problem
    Shared files grow quietly; eventually agents mostly reread their own history.
    Trade-off
    Agents lose sight of anything older than the current sprint.
    Mechanism
    Size caps plus archive rules, and a build gate that fails on growth.

  4. The system treats my attention as the scarcest resource.

    framework — 2026-06-07
    Decision
    Runs halt at planned gates — never interrupt, never guess.
    Problem
    Unattended agents guess wrong or stall — constant pings don't scale.
    Trade-off
    No steering mid-run — ambiguity is paid up front, at planning.
    Mechanism
    The run doesn't need me present — it waits only at gates, for a written verdict.

§05 · SHIPPED WITH MUSTER

Bodh — a shipped iOS app and web landing, live at bodh.day. The replay above is its website wave.

Bodh, idea to live: 9.3 hours of active build, $147 in AI tokens at API list price.

Muster was extracted mid-build from a real iOS app — the framework existed as working practice before it existed as a repo.

This page — built by its own AI team, and measured with the same meter.

BODH · IDEA → LIVE

OPERATOR ATTENTION
4.8 h
SHIPPED
bodh.day
App Store + web

THIS SITE · SPEC → LIVE

OPERATOR ATTENTION
7.5 h
SHIPPED
THIS PAGE

§06 · GET STARTED

One command. No signup, no framework install, no API wiring — markdown files and Claude Code. The gates are deterministic bash — the framework's correctness doesn't depend on which model runs it.

curl -fsSL https://raw.githubusercontent.com/thinkArhant/muster-ai/main/scripts/setup-project.sh | bash -s my-product

cd my-product && claude

github.com/thinkArhant/muster-ai