🎼 Claude Orchestra

A routing & governance layer for your Claude Code skills, agents & MCPs.

I went from zero coding to 500+ Claude Code skills, agents, and MCP servers installed in a few months. At some point I had no idea what was active, what would fire for a given task, or whether I'd installed the same thing three times. More tools had made me slower. Claude Orchestra is what I built to fix that.

Built to be forked. This isn't a one-click install you forget about — it's a template you fork, customize with your rosters, and run on your own machine.

The problem

WITHOUT Claude Orchestra:

  ~/.claude/ → 247 skills · 40 agents · 12 MCPs · 8 plugins
  → "Which one handles this task?"            🤷
  → "Did I install something for this already?" 🤷
  → "Why did THAT skill just fire?"            🤷

WITH Claude Orchestra:

  → architect: BUILD stack — code-reviewer, debugger, gh
  [the work continues here]

Every request routes through your constitution — silently. The only ceremony left is a one-line announcement when work is delegated to a specialist (v2's always-on 🎼 ensemble blocks are gone in v3; the work leads the reply, not the routing).

The fix: 22 orchestras

Claude Orchestra files every tool into themed orchestras. Each has one conductor that sequences its players, clear triggers, and quality gates. A routing hook reads every request and activates the right orchestra(s) automatically.

Compound requests stack:

→ design-ux-architect: landing-page mockups, then handoff to architect for build
  (DESIGN → BUILD stack, sequenced by the constitution)

Idea / business-planning prompts fire NEXUS, the meta-conductor that sequences whole orchestras across a 7-phase lifecycle:

→ NEXUS: Phase 0 Discovery — sequencing RESEARCH · PRODUCT · KNOWLEDGE

The 22 orchestras

The example config (examples/my-22-orchestras.md) shows a real filled-in setup — rosters, conductors, and the reasoning behind each placement. Every skill links to its original author. Use it as a starting point and prune to your own stack.

What's inside v3

Seven doctrine upgrades shipped in v3 (all in orchestra-system.md):

• 22 orchestras — ㉑ AUDIT (quality gate, defaults NEEDS WORK) and ㉒ CYBERSECURITY (a slot for your security tools) added
• Silent routing — v2's always-on 🎼 ensemble announcements are gone. Routing happens quietly; the only ceremony is a one-line delegation announcement (→ agent: task) when work is actually handed off.
• Gap → Recommend — when nothing installed fits a task, the doctrine searches the skill registries and presents vetted install candidates (installs, source reputation) instead of improvising with a generic approach. Recurring gaps are install signals.
• Internal-first search ladder — before going to the web, the doctrine has the model check internal sources (your notes/knowledge base → ~/.claude docs → session memory) first, and only then search externally. Saves tokens, finds prior work first.
• Harmony — 3-Layer Ensemble — each orchestra activates the best combination of tools (an Active Ensemble loaded in full + a Standby Bench summoned by name + a named-only Reserve Bench), not just the single top match.
• Calibration loop — surprising failures get logged to an append-only learnings file and anchored into the doctrine, so the same mistake can't recur.
• NEXUS phase map — idea/business-planning prompts sequence orchestras across a full 7-phase lifecycle (Discovery → Strategy → Foundation → Build → Hardening → Launch → Operate), not just activate one team.

These are doctrine, not enforced code. The hook injects the routing directive and the model honors it — strong conventions, not a runtime engine. Tune them to your taste.

Optional power-up (not installed by default): if you run a local semantic index over your skill catalog (e.g. qmd or any embeddings store), the skill-selector skill can score every installed skill against each prompt instead of relying on lexical matching. This needs extra tooling (Python + an index) — the core system stays zero-dependency without it.

How NEXUS sequences orchestras

Cross-cutting: Superpowers

Some skills apply across ALL orchestras regardless of which fired. These run on discipline, not routing:

Requirements

That's it. No Python, no npm, no Docker.

Install — the audit-first path (recommended)

# 1. Clone
git clone https://github.com/Momo2323-ui/claude-orchestra
cd claude-orchestra

# 2. Audit — read every line you're about to run
cat install.sh

# 3. Preview — see every action without touching anything
./install.sh --dry-run

# 4. Install — idempotent, backs up every file before it changes it
./install.sh        # or: ./install.sh --minimal   (just the load-bearing trio)

# 5. Verify what changed
diff "$(ls -t ~/.claude/settings.json.bak.* | head -1)" ~/.claude/settings.json
ls -la ~/.claude/skills/orchestra-router ~/.claude/agents/auditor.md

# 6. Customize — edit your roster
$EDITOR ~/.claude/rules/orchestra-system.md

Open a new Claude Code session — the router fires on your next prompt.

Install — the fast path

If you already trust the repo:

curl -fsSL https://raw.githubusercontent.com/Momo2323-ui/claude-orchestra/main/install.sh | bash

Or let Claude Code drive it:

Install this for me: https://github.com/Momo2323-ui/claude-orchestra

Want to reverse the install? See UNINSTALL.md.

What the installer actually does

These are Claude Orchestra's own routing/governance skills + the auditor agent — no third-party tools are bundled. Everything backs up before it mutates. Run ./install.sh --dry-run to see every action without touching the filesystem, or --minimal for just the load-bearing trio.

No network calls (except the optional curl-pipe clone), no sudo, no telemetry. Full breakdown in SECURITY.md.

How it works

1. The constitution (orchestra-system.md) defines your orchestras — rosters, conductors, triggers, gates.
2. The routing hook injects a routing directive on every prompt.
3. The router skill matches your request to the right orchestra(s) and announces them.
4. The intake skill files anything new you install into the right orchestra — never archived.

→ Full walkthrough in docs/HOW-IT-WORKS.md.

Build your own orchestras

docs/CREATE-YOUR-ORCHESTRA.md walks you through mapping your tools into orchestras in about ten minutes.

See a real setup

examples/my-22-orchestras.md is a real 22-orchestra config — rosters, the reasoning behind each placement, and links to every skill so you can install the ones you like.

FAQ

Does this install any skills/agents for me? Only its own — the routing/governance skills (orchestra-router, orchestra-intake, skill-selector, hallucination-guard) and the auditor agent. It does not bundle anyone else's tools. The example config links to third-party skills at their original repos so you install those yourself (and the authors get the credit).

Will it overwrite my settings? No. The installer backs up settings.json, merges with jq without clobbering existing hooks, and only appends to CLAUDE.md if the rule isn't already there.

Is this on PyPI / npm / homebrew? No. It's a Bash + Markdown repo. Clone it, audit it, run the installer. Keeping it small is the point.

Do I have to use all 22 orchestras? No. Delete, merge, or rename freely. The themes are a starting taxonomy, not a rulebook.

Does the curl installer give me the orchestras? The curl path installs the infrastructure (hook + router + intake skills). The 22-orchestra roster is a template you customize — edit ~/.claude/rules/orchestra-system.md to map your own tools in.

Do I need NEXUS? No — it's optional. Delete the section if you don't want a meta-conductor.

Can I uninstall? Yes — see UNINSTALL.md.

Security

Every install modifies your ~/.claude/. We take that seriously — see SECURITY.md for the full breakdown of what gets touched, how to audit before running, how to verify after, and how to report a vulnerability privately.

Contributing

PRs and new-orchestra ideas welcome — see CONTRIBUTING.md.

A note for AI-assisted contributors: please verify every command, package name, and feature in your PR actually exists in this repo before submitting. We've had AI-drafted PRs invent commands and features that don't exist — those will be closed.

Credits

Built by Moksh Mittra. MIT licensed — use it freely. Every skill referenced in the example links to its original author.