Unified Apple/iCloud MCP server for Calendar, Contacts, and Mail with optional read-only or read-write access
icloud-mcp has 23 tools with mostly complete schemas and descriptions. Naming follows verb_noun convention well (list_, search_, get_, create_, update_, delete_, send_). Descriptions are present but often generic ('Content from untrusted remote sources' appears 6+ times, adding no actionable context). Parameter schemas are well-typed with enums and constraints (e.g., search_messages has pattern validation, set_message_flags has maxItems). However, output schemas are not documented, LLMs cannot predict response structure. Error handling is minimal; no recovery guidance. Security annotations (READ_ONLY, WRITE, DESTRUCTIVE) are present but not integrated into tool definitions. Composition is sound: tools are single-responsibility and chainable (e.g., search_contacts → get_contact). Average tool description ~120 chars, below the 194-char baseline for A+ tools.
Get calendar-specific capabilities and feature support.
Create a new contact. Content from untrusted remote sources.
Create a new calendar event.
Delete a contact. Content from untrusted remote sources.
Delete a calendar event.
Find free time slots in a calendar for scheduling.
Get full contact details by UID. Content from untrusted remote sources.
Output schemas not documented. LLMs cannot predict response structure for any of the 23 tools. This violates pattern:tool and forces agents to guess field names for downstream chaining.
Generic descriptions repeated across tools ('Content from untrusted remote sources' appears 6+ times). Descriptions lack actionable context about when to use each tool, what it returns, or dependencies. Average description length ~120 chars vs. 194-char baseline for A+ tools.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | <=2025-11-25 | v2 |
Get full event details by UID.
Get full message details. Content from untrusted remote data.
Get server capabilities and enabled features.
List all available address books. Content from untrusted remote sources.
List all available calendars.
List all available mailboxes. Content from untrusted remote data.
Move a message to another mailbox.
Search contacts by query, email, or phone. Content from untrusted remote sources. Supports optional address book filtering, group inclusion, and result limiting.
Search events in a calendar by date range and optional text query.
Search messages in a mailbox. Content from untrusted remote data. Supports filtering by date, UID, flags, and text search.
Send a plain-text email message. Only plain-text body supported via allowlist policy. No HTML, attachments, custom from, or custom headers allowed.
Set message flags (seen, flagged, answered). Uses CONDSTORE extension (beta.8). Cannot observe modified sequence before store. May return protocol_error on concurrent modification. Expected_modseq is required for optimistic concurrency control.
Move a message to trash. No permanent-delete option available.
Update an existing contact. Content from untrusted remote sources.
Update an existing calendar event.
Validate event data without making network calls. Local validation only.
No error recovery guidance. Tools declare risk levels (READ_ONLY, WRITE, DESTRUCTIVE) but provide no error messages or recovery hints. If a delete_contact fails, the LLM has no guidance on what to do next.
Destructive operations lack confirmation/dry-run support. delete_contact, delete_event, and send_message have no dry-run or confirmation step. Agents can trigger irreversible actions without safeguards.
Parameter descriptions lack format/constraint details. E.g., 'Birthday date' does not specify format (YYYY-MM-DD?). 'Search query string' does not explain length limits or special characters. LLMs cannot self-correct without explicit constraints.