MCP tools

Quartermaster exposes its whole surface over an MCP Streamable-HTTP server — 38 tools in ten areas.

Discover & search

Tool	Purpose	Writes?
`search_capability(query, limit?, type?, tool?)`	Hybrid BM25 + semantic search across every indexed type/tool; duplicates collapsed (one result, all locations noted)	read-only
`fetch_capability(id\|slug\|hash)`	Full body + canonical manifest + snapshot for a match	read-only
`list_capabilities(type?, tool?, limit?)`	Browse the inventory with filters	read-only
`census()`	Total vs unique counts, breakdowns by type/tool/scope	read-only
`reindex()`	Re-crawl the scan roots (incremental by content hash)	index only

Tool	Purpose	Writes?
`assess_compatibility`	Can this item hydrate to a target tool? Returns `native` / `convertible` / `blocked` with confidence, warnings, and required conversions	read-only
`convert_capability({ item, target })`	Dry-run conversion preview to another tool, with `warnings`, `lossiness`, and `blockers`	preview

Tool	Purpose	Writes?
`preview_hydration`	Draft an install plan (operations marked `dryRun: true`)	preview
`apply_hydration`	Execute an approved plan — rejects high-risk, unknown-origin, blocked, or policy-violating plans; every target path validated against `HYDRATION_ROOTS`	apply
`rollback_hydration`	Restore recorded files and remove only what the apply created	apply (reversal)

Tool	Purpose	Writes?
`plan_skill_improvement({ method, itemIds \| all })`	Dry-run improvement plan for one, several, or all indexed skills — verifies engine prerequisites (SkillOpt / SkillOpt-Sleep / darwin-skill), returns warnings plus step-by-step engine instructions; never runs the engine	preview
`apply_skill_improvement({ itemId, method, improvedBody })`	Write a reviewed improved body over the live `SKILL.md` — requires `approval: "approve"`, root-confined with a pre-write backup, records an audit row	apply
`list_improvement_history(itemId?)`	The audit trail: method, before/after hashes, status, timestamps, notes	read-only
`rollback_improvement({ runId })`	Restore the recorded pre-improvement content (once per run) — requires `approval: "approve"`	apply (reversal)

Tool	Purpose	Writes?
`export_pack`	Build a v1 (bodies) or v2 (registry-backed, with provenance/risk/drift/compatibility sections) pack manifest	preview
`import_pack`	Preview an import; every artifact marked `trustState: "untrusted"`, nothing hydrated	preview
`verify_pack`	Recompute and check every artifact hash	read-only
`materialize_pack`	Write a v1 pack to disk under `PACK_OUTPUT_ROOTS` — requires `approval: "approve"`, returns a rollback handle	apply

Tool	Purpose	Writes?
`register_device`	Register a device profile (rows in the local DB only)	metadata
`list_devices`	List registered device profiles	read-only
`census_devices`	Per-device inventory census	read-only

Tool	Purpose	Writes?
`scan_risk`	Deterministic local trust scan; high-risk findings block apply	metadata
`list_risk_findings`	List persisted risk findings	read-only
`set_trust_state`	Mark an artifact trusted/untrusted after review	metadata

Tool	Purpose	Writes?
`find_upstream({ itemId })`	Ranked upstream candidates with `matchMethod`, `confidence`, and evidence — sidecar → git origin → package/MCP source → catalog → content fingerprint	read-only
`link_upstream({ itemId, candidate })`	Persist a selected candidate as the artifact’s upstream link	metadata
`pin_version`	Pin the link to a ref/version/commit (suggested config preview for MCP servers — never rewrites config files)	metadata
`unpin_version`	Clear the pin	metadata
`explain_drift({ itemId })`	Full drift status taxonomy, file-level changes, conflicts, semantic findings, risk, recommendation	preview
`scan_drift({ itemIds \| allLinked })`	Compute (and by default persist) drift reports for many linked items	metadata
`accept_local_changes`	Accept a reviewed fork as the new baseline	metadata
`check_updates({ itemId \| all })`	Installed vs latest upstream — cache-only unless the network double-gate is open	read-only*
`preview_update({ itemId })`	The canonical dry-run merge planner: baseline → current → latest → operations	preview

* network only with ALLOW_NETWORK_UPDATES=1 and allowNetwork: true. See Upstream provenance & drift.

Tool	Purpose	Writes?
`refresh_public_catalog({ entries })`	Import static catalog entries (or ingest live from an MCP registry URL behind the network double-gate)	metadata
`list_public_catalog()`	List cached entries	read-only
`match_public_catalog({ itemId \| path \| name })`	Match an item against the cache	read-only

Tool	Purpose	Writes?
`snapshot_mcp_surface`	Store a supplied `tools/prompts/resources` listing (metadata only — never connects to a live server, never executes anything)	metadata
`compare_mcp_surface`	Diff two snapshots; flags added/removed tools and raises a high-severity finding when a new tool has `destructiveHint: true`	read-only

read-only — no state change at all.
metadata — writes rows in Quartermaster’s own SQLite database; never artifact files.
preview — returns a plan (dryRun: true, writes: []); nothing touches the filesystem.
apply — the only tools that write artifact files; approval-gated, root-confined, reversible.

Health/debug endpoints: GET /health, GET / (the web console).