mcp-tools.md

← Back to README

MCP Tool Reference

All 29 tools exposed by the Cortex MCP server. Source of truth: crates/daemon/src/handlers/mcp/.

---

Core

Tool	Required	Optional	What it does
cortex_boot	—	agent, budget, profile	Session boot prompt from identity, delta, and current-truth capsules
cortex_boot_audit	—	agent, limit	Read recent boot audit rows and token/capsule metadata
cortex_peek	query	limit	Headline-only relevance check (~80% cheaper than recall)
cortex_recall	query	budget, policyMode, k, agent, taskClass, adaptive	Clock-Quorum Recall over write, truth, task, and history clocks
cortex_semantic_recall	query	budget, k, agent	Named CQR surface; same engine as cortex_recall, no embedding model
cortex_recall_policy_explain	query	budget, policyMode, k, pool_k, agent	Explain ranking: why these results, in this order
cortex_store	decision	context, type, source_agent, confidence, reasoning_depth	Persist a decision with conflict detection
cortex_unfold	sources	—	Expand memory sources to full text (use after peek)
cortex_health	—	—	System health, DB stats, memory counts
cortex_digest	—	—	Daily summary: activity, savings, top recalls

Memory management

Tool	Required	Optional	What it does
cortex_forget	source	—	Decay matching entries (score × 0.3)
cortex_resolve	keepId, action	supersededId	Resolve a disputed decision pair (keep or merge)
cortex_memory_decay_run	—	includeAging, cleanupExpired	Run maintenance: decay, aging, expired cleanup
cortex_lastCall	—	kind, agent	Fetch most recent memory, decision, or event

Conflicts

Tool	Required	Optional	What it does
cortex_conflicts_list	—	status, classification, conflictId, limit	List conflicts with filters
cortex_conflicts_get	conflictId	—	Fetch one conflict by ID
cortex_conflicts_resolve	action	winnerId, keepId, supersededId, loserId, conflictId, classification, similarity, notes, resolvedBy	Resolve with winner + metadata
cortex_consensus_promote	—	limit, minMargin, dryRun	Auto-resolve when trust margin is clear

Focus sessions

Tool	Required	Optional	What it does
cortex_focus_start	label	agent	Start focus session (context checkpoint)
cortex_focus_end	label	agent	End session, consolidate into summary
cortex_focus_status	—	agent	Active/recent sessions and savings

Agent telemetry

Tool	Required	Optional	What it does
cortex_agent_feedback_record	outcome	agent, taskClass, outcomeScore, qualityScore, latencyMs, retries, tokensUsed, memorySources, notes	Record task outcome with metrics
cortex_agent_feedback_stats	—	horizonDays, limit, taskClass, agent	Reliability trends from outcomes
cortex_eval_run	—	horizonDays	Conflict pressure + resolution snapshot

Permissions

Tool	Required	Optional	What it does
cortex_permissions_list	—	—	List permission grants
cortex_permissions_grant	client, permission	scope	Grant read / write / admin
cortex_permissions_revoke	client, permission	scope	Revoke a grant

Session

Tool	Required	Optional	What it does
cortex_diary	—	accomplished, nextSteps, decisions, pending, knownIssues	Write session state for cross-session continuity
cortex_reconnect	—	agent, model	Re-register after daemon restart or disconnect

---

Quick reference

• Recall budget defaults to 200 tokens. Policy modes: fast, balanced, deep.
• Conflict classes: AGREES, CONTRADICTS, REFINES, UNRELATED.
• Feedback outcomes: success, partial, failure.
• Permission levels: read, write, admin. Default scope: *.
• Progressive disclosure: peek (headlines) → unfold (full text of selected items) → recall (full search).