Inspect and debug CloudKit databases, records, zones, and sharing through a secure, read-only MCP server.
CloudKit MCP demonstrates solid definition quality with consistent naming patterns, comprehensive parameter schemas using Zod validation, and clear descriptions for most tools. All 12 tools follow verb_noun naming (get_, list_, query_, etc.). Schemas are well-structured with proper type constraints, enums, and length limits. However, parameter descriptions lack depth, most are generic ('View configuration', 'Zone identifier') without explaining WHEN to use each tool or HOW they differ. Output schemas are not documented in the tool definitions themselves. Error handling descriptions are absent. The server correctly avoids exposing secrets and implements read-only operations, but lacks guidance on recovery paths and error categorization.
Compare exact record observations made with two independently configured views and zone selectors.
Report offline profile policy and capability state without loading credentials or contacting Apple.
Read bounded database change state from an explicit beginning or process-bound cursor.
Look up metadata for at most twenty exact records in one owner-aware zone.
Inspect bounded privacy-safe share topology attached to a proven record.
Look up one exact owner-aware zone.
Parameter descriptions are generic and lack context. 'View configuration with profileId and scope' does not explain WHEN to use public vs private vs shared scope, or what each scope reveals. LLMs cannot infer scope selection logic.
Output schemas are not documented in tool definitions. Callers cannot know what fields to expect (e.g., does query_records return record IDs, metadata, or full payloads?). This forces LLMs to guess downstream field names.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2025-06-18+ | v2 |
Read bounded custom-zone changes and tombstones from an explicit beginning or process-bound cursor.
List supported subscription structure without notification payloads or predicate values.
List bounded owner-aware zones for an explicit view in an enabled documented scope.
Perform one minimal authentication/current-principal probe; an API-token-public profile may return the documented user-authentication challenge without database access.
Run one bounded typed indexed query; arbitrary predicates and cross-zone scans are unavailable.
Read only exact payload fields enabled by immutable startup policy.
No error handling guidance. Tools lack descriptions of failure modes (e.g., 'If profileId is invalid, returns 404 with available profiles'). Agents cannot self-correct without recovery hints.
Tool descriptions do not explain tool selection logic. 'get_records' vs 'query_records' distinction is unclear, when should an LLM choose one over the other? Descriptions should state prerequisites and use cases.
Pagination support is incomplete. query_records accepts continuationHandle but does not document what the response contains (does it include nextContinuationHandle?). Agents cannot implement pagination loops without guessing response structure.