Metro-North MCP has 11 well-named tools with consistent verb_noun patterns (get_*, search_*, plan_*). All tools have descriptions (avg ~80 chars, within baseline 34-392 range). Input schemas are present and typed for all tools. However, descriptions are terse and lack context for LLM selection, most are single-sentence statements without WHEN/WHY guidance. Parameter descriptions exist but are minimal (avg ~40 chars vs baseline 72). Output schemas are not documented in the source. Error handling and recovery guidance are absent. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all tools being READ_ONLY. Composition is sound, tools are single-responsibility and chainable (e.g., search_stations → get_station_info → get_departures). Security posture is good: no secrets in parameters, all tools are read-only, no destructive operations.

• Expand tool descriptions to 100-200 chars. Include WHEN to use (e.g., 'Call this to see real-time departures for a specific station. Use get_route_schedule for the full day's timetable.'), WHAT it returns (e.g., 'Returns a list of upcoming trains with departure times and current delays'), and any prerequisites (e.g., 'Requires a valid station name, use search_stations first if unsure').
• Document output schemas for all tools. Add a 'Returns' section to each tool description listing the fields in the response (e.g., 'Returns: {trip_id, departure_time, arrival_time, stops[], delays_minutes}').
• Add readOnlyHint: true to all tool definitions in the MCP server code (src/tools/handlers.ts). This signals to LLMs that these tools are safe to call without confirmation.
• Enhance parameter descriptions with format and constraint details. E.g., 'direction: One of inbound (toward Grand Central) or outbound (away from Grand Central). Defaults to inbound if omitted.' and 'limit: Maximum results to return (1-100, default 20).'
• Add error recovery hints to tool descriptions. E.g., 'If station_name is not found, call search_stations() with a partial name to discover valid station names.' This guides LLMs to self-correct on invalid input.

Inferred effective spec: 2026-07-28+.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry