Configuration

Muster needs no config file. Every tunable is an environment variable, every one has a working default, and every one is read at the point of use — so exporting it in your shell, your settings.json env block, or a single command invocation all work the same way.

Integer variables are parsed strictly: a non-numeric or out-of-range value falls back to the default rather than failing the run. Boolean-shaped declarations (MUSTER_ENABLE_APEX, MUSTER_AGENT_TEAMS, MUSTER_COWORK_NATIVE_PLUGIN) accept 1/true to enable and 0/false to disable; anything else fails closed to the documented default.

Variables

Variable	Default	Semantics	Read by
MUSTER_INLINE_SCALE	3	Border-invitation threshold: the Nth distinct file edited inline across turns, with no muster run active, warns once per crossing. Never denies. Minimum 2.	PreToolUse hook (plugin/hooks/inline-budget.js)
MUSTER_INVITE_COOLDOWN_MS	900000 (15 min)	Suppresses a repeat border invitation for this long after the last one actually fired, so a session is never nagged twice in quick succession.	UserPromptSubmit + PreToolUse hooks
MUSTER_ACTION_GUARD	deny	The action-class fence, active only while a run is live and .muster/forbidden-actions lists a class: deny blocks a matching send/sign/submit/publish/purchase/delete-remote tool call, warn allows it with a reminder, off disables the fence. The only hard-deny surface in the stack.	PreToolUse hook (plugin/hooks/pre-tool-use.js)
MUSTER_TASK_GATE	(unset — gate on)	Ties a native task board "completed" tick to a recorded review-gate PASS for muster-tracked tasks; a pending or escalated entry is denied. off allows unconditionally. Fail-open for any task muster did not create.	TaskCompleted hook (plugin/hooks/task-completed-gate.js)
MUSTER_MAX_TIER	(unset)	Caps the conceptual model tier policy (prime excludes apex, core enables budget mode); unset means no cap. Static agent frontmatter pins are unaffected on direct invocation; in Muster runs the dispatch override honors the cap.	CLI (src/model.js)
MUSTER_ENABLE_APEX	(unset; off)	Enables apex for peak-judgment roles (tournament judge, architecture-review, improve, advisor). Unset degrades apex to prime deterministically.	CLI (src/model.js)
MUSTER_ADVISOR_MAX_CONSULTS	3	Maximum advisor consults per run — bounds the cost of workers escalating to the advisor role. 0 disables advisor consults.	CLI (src/advisor.js)
MUSTER_FUSE_TOPK	3	Maximum tournament candidates passed to the fusion synthesizer. Minimum 1.	CLI (src/fusion.js)
MUSTER_FUSE_MIN_DISAGREEMENT	1	Minimum disagreement score required to activate fusion synthesis; below it, muster fuse falls back to the single best candidate. 0 always fuses when at least two candidates pass.	CLI (src/fusion.js)
MUSTER_REVIEW_DIFF_THRESHOLD	200	Changed-line threshold at which the review gate dispatches two reviewers instead of one. A batching lever only — the gate's pass bar and fix-loop cap are unchanged.	CLI (gate-cadence, src/gate-cadence.js)
MUSTER_AGENT_TEAMS	(unset — off)	Declares that this session's tool list carries the native Workflow fan-out/barrier surface, selecting mode: "native" wave dispatch. Nothing declared means the prose wave loop, the floor on every harness. An explicit --agent-teams/--no-agent-teams flag always wins over the env var.	CLI (wave-dispatch, src/wave-dispatch.js)
MUSTER_SPRINT_PARALLEL	5	Max concurrent item-runner subagents per wave in /muster:go-backlog wave mode. Hard ceiling 10 (higher values clamp; 0 is invalid). Read by go-backlog's orchestration protocol, not by library code.	plugin/commands/go-backlog.md

Harness-scoped variables

These are read only on the harness they name, and mostly by its host rather than by you. The Cowork adapter is cowork/mcp-server.mjs; the neutral core receives its runtime identity explicitly rather than from a user-set host selector.

Variable	Default	Semantics	Read by
MUSTER_RUNTIME	(unset)	The neutral MCP core marks nested CLI invocations with the active host (cowork, codex, or work), so capability resolution uses that lane. Set on child CLI spawns by mcp/server.mjs.	CLI (src/capabilities.js)
MUSTER_COWORK_NATIVE_PLUGIN	(unset — off)	Declares that Cowork's own plugin loader accepted muster's plugin/ tree. A declared capability check, never a probe: unset keeps resolution MCP-only.	CLI (src/cli.js, src/capabilities.js)
MUSTER_COWORK_CONNECTORS	(unset)	Comma-separated remote-connector names (e.g. slack,drive) to treat as available. Remote connectors live in your cloud account, not on disk, so they cannot be auto-discovered.	CLI (src/cli.js)
MUSTER_COWORK_MAX_INFLIGHT	4	Max concurrent MCP tool executions in the Cowork server. Hard ceiling 64.	mcp/server.mjs (via cowork/mcp-server.mjs adapter)
MUSTER_COWORK_MAX_QUEUE	16	Max queued MCP tool executions before overload rejection. Hard ceiling 1024.	mcp/server.mjs (via cowork/mcp-server.mjs adapter)

Kimi also reads host-owned settings rather than MUSTER_* settings. KIMI_CODE_HOME selects the Kimi data root. KIMI_SECONDARY_MODEL and KIMI_CODE_EXPERIMENTAL_FLAG enable the per-process secondary model lane used by Muster-launched Kimi processes. Kimi's own loop and background concurrency variables remain Kimi runtime controls, not Muster configuration. See the Kimi guide before changing them.

Where to set them

Every variable is read at the point of use, so all three of these work and none of them need a restart of anything but the process that reads the value:

# one invocation
MUSTER_MAX_TIER=core npx -y @adnova-group/muster capabilities

# this shell, this session
export MUSTER_ENABLE_APEX=1

For the hook-side variables (MUSTER_INLINE_SCALE, MUSTER_INVITE_COOLDOWN_MS, MUSTER_ACTION_GUARD, MUSTER_TASK_GATE), set them where the harness can see them when it spawns a hook — the env block of your Claude Code settings.json, or the environment Codex itself was launched from. A variable exported inside a session after the hook has already fired does not retroactively change that firing.

To confirm a value is actually reaching the CLI, read it back off a command that reports it: muster capabilities shows the effective per-role model (proving MUSTER_MAX_TIER and MUSTER_ENABLE_APEX), and muster gate-cadence <manifest.json> --changed-lines N shows the resolved reviewerCount (proving MUSTER_REVIEW_DIFF_THRESHOLD).

Muster's conceptual ladder is scout, core, prime, and apex; harness adapters map those tiers to concrete models.

Compatibility only: legacy input aliases haiku, sonnet, opus, and fable normalize to scout, core, prime, and apex; they are not a second conceptual ladder. MUSTER_ENABLE_FABLE remains an alias for MUSTER_ENABLE_APEX.

Not user knobs

Three more MUSTER_* names appear in the tree and are deliberately absent from the tables above:

• MUSTER_CLI is a run-local shell variable, not a setting. Each mode resolves the cheapest way to invoke the CLI once per run (a vendored plugin runtime, a local checkout, a muster bin, or an npx fallback) and reuses the answer, so a run never pays a cold npx -y start per call. Use muster resolve-cli to see what it would resolve to.
• MUSTER_TEST_NOW_MS and MUSTER_COWORK_TEST_CLI are test-only injection points (a frozen clock for the hook cooldown tests; a CLI path override honored only under NODE_ENV=test). Neither affects a real run.

Next: Troubleshooting for what to do when a setting is not taking effect.

Harness setup and support boundaries are documented in Harness support, Codex, Kimi, and Cowork. Security-sensitive diagnostics and reporting use Security and reporting.