Install

Muster has two parts that install together: a deterministic CLI (an npm package) and a Claude Code plugin. The CLI does the routing math; the plugin teaches Claude Code how to drive a run.

Requirements

• Claude Code
• Node.js 20 or newer (node --version)

Model work uses the account or subscription of the active Claude Code session. Muster's CLI makes no model calls, so there is no separate model API, hosted runtime, or key to manage.

1. Run the installer

npx -y @adnova-group/muster@0.6.0 install

install mutates nothing in your ~/.claude. It only prints the steps it cannot do for you, because registering a plugin is a Claude Code action, not a shell command.

2. Register the plugin

Run these inside Claude Code:

/plugin marketplace add Adnova-Group/muster   # register the marketplace
/plugin install muster@muster                 # install the plugin

Muster's glass-box output style ships inside the plugin and applies automatically when the plugin is enabled (it sets force-for-plugin: true), so there is no style command to run. The old /output-style <name> command was removed from Claude Code in v2.1.91; auto-apply replaces it. To pick a different style at any time, use /config and select Output style.

Restart to activate

Plugin install is a Claude Code action, so the running session only picks Muster up after you (re)install it through /plugin. The plugin's agents, the four session hooks, and the output style become active in your next fresh session (restart or /clear).

Using another runtime?

Codex, Kimi, and Cowork have their own install and enforcement paths. Start with Harness support.

3. Verify

Start a new Claude Code session and run:

/muster:plan Add rate limiting to the public API with tests

Muster detects your project, assembles a crew, and shows the glass-box manifest plus a plan, then stops for your approval. If you see the crew manifest, you are set.

You can also exercise the CLI directly in a terminal. Every verb is plain Node and prints JSON:

npx -y @adnova-group/muster detect
npx -y @adnova-group/muster capabilities

If the crew manifest does not appear, or /muster:* commands are missing, run npx -y @adnova-group/muster doctor and see Troubleshooting.

What the plugin adds

Claude Code exposes all ten modes, including Design and Init.

• Ten slash-command modes: /muster:plan, /muster:go, /muster:plan-backlog, /muster:go-backlog, /muster:diagnose, /muster:audit, /muster:design, /muster:runner, /muster:capture, and /muster:init. /muster:run, /muster:autopilot, and /muster:sprint still work as aliases of plan, go, and go-backlog. Deprecated as of 2026-07-17 and retiring in muster 0.7.0; behavior stays unchanged until then.
• Four session hooks, all declared in plugin/hooks/hooks.json and active only while Muster is enabled. Enforcement follows the run's EXTERNAL effects, not the orchestrator's own in-repo edits: the action-class fence below is the only hard deny on a tool call, and the TaskCompleted gate below is the one other block surface (it gates a task-board completion tick, not a tool call); everything else is a single warn-only "border invitation" that sells the value of a crew run rather than commanding.
  • SessionStart injects a one-line pointer ("muster available; /muster:plan for orchestration-scale work") into every session, and clears stale run/session state on a genuinely fresh start. Never writes to your ~/.claude files.
  • UserPromptSubmit fires the ONLY prompt-time nudge: a directive-shaped prompt (fix/build/implement, etc.) with no active run sells the value of a crew run (parallel dispatch, adversarial review, a receipts trail) once per crossing, then stays silent until a run starts, a fresh session, or 60 minutes of inactivity re-arms it.
  • PreToolUse enforces the action-class fence: while a run is active and .muster/forbidden-actions lists a class, every payload is classified before exemptions and a tool call that would perform a run-forbidden send/sign/submit/publish/purchase/delete-remote action is denied, controlled by MUSTER_ACTION_GUARD (deny / warn / off). Independently, a cumulative counter of distinct inline-edited files across turns (with no run active) crossing MUSTER_INLINE_SCALE (default 3) warns once per crossing with the same value-toned copy -- never denies. Writes into .muster/ and .claude/ (in-cwd repo) are exempt only from that advisory drift accounting.
  • TaskCompleted is the second block surface. Muster writes .muster/task-board.json (one entry per native task it created) and flips an entry to reviewGate: "pass" only once the review gate actually passes that task. The hook denies (exit 2) a completion tick on a tracked task with no recorded PASS, so a task cannot be ticked done before it has been reviewed. If you see a completion refused mid-run, that is this gate: finish the review, or set MUSTER_TASK_GATE=off to disable it. It fails open for any task the board does not track.
• Built-in agents and skills, vendored from MIT-licensed upstreams plus Muster's own clean-room specialists.

Uninstall

Because everything Muster adds lives inside the plugin, removal is mostly a matter of removing the plugin. Muster never writes to your ~/.claude/CLAUDE.md or settings.json, so there is nothing tangled to unpick.

npx -y @adnova-group/muster@0.6.0 uninstall

uninstall prints the steps it cannot do for you, because removing a plugin is a Claude Code action:

/plugin uninstall muster@muster    # remove the plugin (and its style + hook)
/plugin marketplace remove muster  # remove the marketplace

It also cleans up after older Muster versions: if a pre-force-for-plugin install left a copied style at ~/.claude/output-styles/muster.md, uninstall removes it (and restores the original it had displaced, if there is a .bak). On a current install there is nothing there to remove.

Everything leaves with the plugin

The output style (force-for-plugin) and all four session hooks are plugin-native, so uninstalling the plugin removes them automatically. The forced style auto-reverts to whatever output style you had before. There is no global file or CLAUDE.md block to clean up by hand.

Next: the Quickstart.