Reference

MCP tools and resources

Run knowl serve to expose Knowl over stdio MCP. The recommended agent flow is:

  1. Use lifecycle bootstrap context when available; otherwise use knowl_recent.
  2. Call knowl_query before inspecting repository files, using the words that name the subject. Another on-subject term retrieves better and an off-subject one retrieves worse, so do not pad a query to reach a length and do not trim a real term to shorten it.
  3. Verify misses, conflicts, or stale results against the repository.
  4. Store durable findings, stated goals, and recurring diagnoses, and update contradicted memory promptly.
  5. Use manual task tools only when verified lifecycle hooks are unavailable.

Tools

Knowl exposes the core tools below. Two transcript search tools and a session listing tool are registered in addition when transcript indexing is enabled for the repository.

Tool Purpose
knowl_query Focused retrieval before files and before each new subtask or project area
knowl_recent Compact recent context when lifecycle bootstrap is unavailable or a refresh is needed
knowl_state Broad active-memory status or hierarchical project summary
knowl_context Compose an explicitly token-budgeted local context pack
knowl_task_start Start one manual work loop when verified lifecycle hooks are unavailable
knowl_task_checkpoint Checkpoint meaningful manual-loop progress or blockers
knowl_task_finish Finish one manual work loop after verification
knowl_store Store one concise structured atom
knowl_ingest_atoms Batch-store client-extracted atoms
knowl_decide Record a confirmed decision and reasoning
knowl_update Correct or supersede stale or contradicted memory
knowl_timeline Inspect one item's immutable assertion history
knowl_evidence_list Inspect evidence linked to one item
knowl_conflicts Inspect active exclusive conflict identities
knowl_feedback Record usefulness or correction feedback after an item is used
knowl_skill_list List learned file-backed skills
knowl_skill_read Inspect one learned skill before running it
knowl_skill_run Run a trusted learned-skill entrypoint
knowl_skill_create Create a learned skill when explicitly requested
knowl_ingest Process explicitly supplied raw source through configured AI
knowl_synthesize Create or refresh one explicitly tag-scoped understanding
knowl_session_finish Finish an explicitly owned manual memory session
knowl_gc_preview Preview duplicate, stale, or cold-memory maintenance
knowl_gc_apply Apply previewed maintenance after explicit approval
knowl_handoff Park a workstream for the next session in this project, delivered once
knowl_park Park a workstream under a short key the user keeps
knowl_resume Resume a parked workstream from its key, from any directory

There is no MCP ask tool. An MCP client model can query structured results directly; raw-source processing is the separate provider-backed knowl_ingest tool.

Resources

Resource discovery advertises only:

Resource Purpose
knowl://brain Local active project memory, formatted and capped at 3,000 characters
knowl://recent Compact local recent session and project context

knowl://category/<name> is directly readable for active items in a category, but it is not advertised by discovery and there is no resource template. These resources are local views, not workspace-federated views; knowl://brain is bounded rather than a complete database dump.

Auth: Unsupported is expected for this local stdio server and does not indicate that the tools are unavailable.