Manage Google Tasks from MCP-compatible AI clients.
Google Tasks MCP demonstrates solid tool design with 19 well-named, verb-prefixed tools (list_, create_, get_, update_, delete_, search_, add, complete, uncomplete, move, clear_, digest, today, overdue, upcoming). All tools have descriptions (avg ~100 chars, within 10-1024 baseline). Input schemas are present and typed for all tools. However, output schemas are not documented in the source code, LLMs cannot predict response structure. Parameter descriptions are present but terse (avg ~40 chars vs 72 baseline). Error handling is minimal; no recovery guidance visible. Tool composition is strong: tools chain well (e.g., list_tasklists → get_tasklist → add task), and natural identifiers (tasklist title, task title) are accepted alongside IDs, reducing lookup friction. Idempotency metadata is declared in _ToolMetadata but not exposed to the MCP protocol via tool annotations.
Create a task or subtask in one tasklist, optionally after a sibling.
Hide completed tasks in one tasklist after confirm=true and report the cleared count.
Mark one task complete by ID or exact title and return a rich mutation response.
Create a tasklist and return compact metadata with a human summary.
Delete one task by ID or exact title and return pre-deletion details.
Delete a tasklist by ID after confirm=true; non-empty lists require force=true.
Return a short text digest; omitted tasklist summarizes all tasklists.
Output schemas not documented. LLMs cannot predict response structure, field names, or types. Responses must be inferred from tool behavior, increasing hallucination risk and forcing agents to guess at downstream field names.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) not exposed via MCP protocol. Metadata exists in _ToolMetadata dict but is not passed to FastMCP tool registration, so clients cannot distinguish read-only from destructive operations.
Error handling lacks recovery guidance. No visible try-catch blocks or error messages that tell LLMs what to do next (e.g., 'User not found. Try search_users() first'). Errors likely return raw exceptions without actionable context.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
Get one task by ID or exact title from one tasklist, including notes and link details.
Get one tasklist by ID, or by exact title when ID is omitted.
List compact tasklist IDs and titles for the authenticated Google account.
List tasks from one tasklist with filters and pagination; omitted tasklist uses the default list.
Move one task by ID or exact title, optionally changing tasklist, parent, or sibling order.
List incomplete overdue tasks; omitted tasklist reads all tasklists.
Search task titles and notes case-insensitively; omitted tasklist searches all tasklists.
List incomplete tasks due today; omitted tasklist reads all tasklists.
Reopen one completed task by ID or exact title.
List incomplete tasks due within N days; omitted tasklist reads all tasklists.
Edit one task by ID, or by exact title for non-title fields.
Rename a tasklist by ID; title lookup is intentionally unsupported.
Parameter descriptions are terse (avg ~40 chars vs 72 baseline). Many lack format/constraint details. E.g., 'due' param says 'Due date in YYYY-MM-DD or RFC 3339 format' but does not specify timezone handling or validation behavior.
Pagination support incomplete. list_tasks accepts max_results and page_token but does not document total count or next_cursor in response. Agents cannot determine if more results exist or how to fetch them reliably.