The server provides 24 tools with consistent naming conventions (verb_noun pattern) and visible tool registration via @mcp.tool() decorators in src/myj_mcp/server.py. However, critical gaps emerge in parameter descriptions, output schema documentation, and error handling. Tool descriptions are present but generic (100-150 chars on average), meeting minimum thresholds but lacking the specificity required for LLM planning. Most tools have parameters with types but NO descriptions, e.g., reserve_equipment has parameters like 'slot_id', 'time', 'duration' that lack per-parameter guidance on format, valid ranges, or constraints. Output schemas are completely undocumented: tools return json.dumps() of model objects, but the actual field structure, cardinality, pagination behavior, and error response formats are invisible. Parameters with 'date' fields lack format validation hints (are partial dates accepted? timezone handling?). No error guidance is provided, a failed authentication or API call will surface a raw exception, not a structured recovery hint. The server relies heavily on _ensure_auth() for state management but does not document retry behavior, timeout handling, or what happens when credentials are invalid. Per-tool scoring reveals consistent weaknesses: naming is solid (average 75), but description and schema scores lag (average 45 each).

• Document output schema for ALL tools. Specify field names, types, cardinality, and whether results are paginated. Example: 'Returns {classes: [{id, name, time, capacity, available_slots, trainer_name}], total_count, next_cursor}'.
• Add per-parameter descriptions to every tool. Use specific format constraints: date must be 'YYYY-MM-DD' (UTC), time must be 'H:MM AM|PM' (local timezone assumed), duration must be 30 - 120 minutes in 15-min increments.
• Implement pagination for list_* and search tools. Add page_size (default 20, max 100) and page (default 1) parameters. Return total_count and next_page hint in response.
• Add error recovery hints. E.g., login tool should document 'If auth fails, check email/password format. If email format is invalid, error_code will be INVALID_EMAIL.' cancel_reservation should return 'If not found, call my_reservations() to confirm the reservation_id.'
• Implement dry-run or confirmation for destructive ops. Add optional 'confirm=true' parameter to cancel_reservation and delete_waitlist. Return a preview before confirming, or require explicit confirmation from the LLM.
• Refactor credential handling. Use a dedicated credential_store tool that accepts email/password via secure input (not environment), or integrate MCP secret injection pattern. Never store plaintext passwords in .env.
• Document idempotency guarantees. Specify which tools can be safely retried (list_*, get_*) and which have side effects (reserve_*, cancel_*, book_*). Add idempotentHint annotations where applicable.