Build secure, typed MCP tools backed by Firestore.
The Firestore MCP Kit presents a well-structured, typed toolkit with clear naming conventions and complete schemas. All 4 tools follow verb_noun naming (notes.create, notes.get, notes.update, notes.delete). Input and output schemas are explicitly defined using Zod and converted to JSON Schema. However, parameter descriptions are minimal (mostly single sentences like 'Note identifier'), and the framework lacks several production-grade patterns: no parameter constraints (enums, ranges, formats), no recovery guidance in error messages, no pagination support (not applicable for single-item operations), and no tool annotations (readOnlyHint, destructiveHint). The authorization pattern for notes.delete is a positive sign for permission gating. Overall, the kit is above average for a community server but lacks the depth of documentation and error handling expected in A-grade tooling.
Create a note.
Delete a note by id.
Get a note by id.
Update a note title and/or body.
Parameter descriptions are minimal and generic. 'Note identifier' and 'Note body, defaults to empty string' lack context about format constraints, length limits, or when/why the parameter is needed. LLMs cannot infer whether 'id' should be alphanumeric, slug-safe, or arbitrary. Per the rubric baseline, A+ tools average 72 chars per parameter annotation; these average ~25 chars.
No input constraints declared in schemas. The 'id' parameter accepts any string (min length 1 enforced in Zod but not visible in JSON Schema output to the LLM). No guidance on format (alphanumeric only? max length?), no enums, no patterns. Per pattern:constrained-input, free-form strings without formal constraints invite hallucinated values.
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 57 | 2025-06-18+ | v2 |
| 2026-03-09 | C | 64 | - | v1 |
No tool annotations (tool destructiveHint, readOnlyHint, idempotentHint). The protocol supports these hints to help agents reason about safety and idempotency. notes.delete should be marked destructive; notes.get should be marked read-only. This information is in the Risk field of the server spec but not exposed in the MCP tool metadata.
Error handling lacks recovery guidance. NotFoundError is thrown but the MCP error response does not suggest recovery actions (e.g., 'Try list_notes to find available notes'). Per pattern:recovery-guide, errors should tell the LLM what to do next.
No idempotency guidance. The notes.create tool will fail if called twice with the same id (Firestore setDocument overwrites, but the tool doesn't declare idempotency). Per pattern:idempotent-operation, agents need to know whether retrying a tool produces the same result or duplicate side effects.
Tool descriptions are below rubric baseline. At 15-23 characters each, they fall well short of the 194-char average for A+ tools. 'Create a note.' and 'Get a note by id.' do not explain when or why to call the tool, what it returns, or dependencies (e.g., does the note id need to be unique? What happens if the id already exists?).