Mind Grapes MCP server — a FastMCP-based brain service for capturing, searching, and managing personal knowledge (experiences, entities, claims) with entity repair, disambiguation, and review workflows.
Strong schema coverage (21 tools with explicit JSON Schema definitions) and comprehensive parameter descriptions. Tool naming follows verb_noun convention well (search_thoughts, capture_thought, merge_entities). Most tools have 50-200 char descriptions aligned with LLM optimization baselines. However, several tools lack output schema documentation, error handling guidance is minimal, and some parameter descriptions could be more actionable. No tool annotations (readOnlyHint/destructiveHint) despite clear risk classifications in metadata. Composition is sound, tools chain well (e.g., resolve_entity → relationships_to). Security concerns: no visible input validation, sanitization, or rate limiting in source.
Capture an image with optional metadata (location, participants, event, OCR, description) and visibility control.
Capture a new experience/thought with optional structured metadata (participants, predicates, source, visibility).
Fetch a single experience by ID with full details (returns found=false if not found or not readable).
List captured experiences with optional filtering by type, topic, person, and recency.
Merge a duplicate entity into a canonical survivor, retiring the loser and updating all references.
Propose a non-destructive correction to an experience, claim, or entity for later review.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite explicit risk classifications in metadata. Tools marked DESTRUCTIVE (merge_entities, retract_claim, split_entity, etc.) should declare this via schema annotations for agent safety.
Output schemas not documented in tool definitions. LLMs cannot plan downstream calls or extract required fields (e.g., what does resolve_entity return? Does it include entity_id for relationships_to?). Breaks tool chaining.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Retrieve recent experiences, optionally filtered by natural-language query, source kind, and time window.
Traverse the entity relationship graph to find connected entities within a hop limit and confidence threshold.
Change the canonical name of an entity.
Request user input to disambiguate between multiple options (e.g., which entity a mention refers to).
Apply or reject a proposed correction, optionally dispatching the suggested change.
Resolve a pending disambiguation by recording the user's choice.
Resolve an entity name to a list of candidate entities, optionally using context for disambiguation.
Mark a claim as retracted (wrong/superseded) without deleting it, preserving audit trail.
Retrieve items awaiting human review (merge candidates, low-confidence claims, contradictions, disambiguations, proposed corrections, split candidates).
Natural-language search over captured experiences with optional filtering by person/topic, confidence thresholds, and provenance attachment.
Split an over-collapsed entity by moving some of its experiences to a new or existing target entity.
Aggregate statistics on captured experiences (counts, top entities, top topics, time range).
Undo a soft merge by clearing the merged_into pointer on an entity.
Update an experience with a patch (idempotent, destructive operations like field deletion).
Find entities (people) present at a specific experience or date.
No error handling guidance in tool descriptions. Tools like merge_entities, split_entity, and resolve_correction can fail (e.g., entity not found, invalid merge, contradictory split). Descriptions lack recovery hints (e.g., 'Try resolve_entity() first to verify the ID').
Destructive operations (merge_entities, retract_claim, split_entity, update_experience) lack confirmation/dry-run support. Agents can irreversibly corrupt the knowledge graph without a safety gate. No visible validation or rollback mechanism.
Parameter descriptions lack actionable constraints. E.g., 'reason' fields say 'audit trail' but don't specify minLength, maxLength, or format. 'threshold' and 'min_confidence' lack range documentation (0 - 1? 0 - 100?). LLMs cannot validate inputs without explicit bounds.