Server defines 8 tools with complete JSON Schema input definitions and descriptions. All tools follow verb_noun naming convention (get_*, find_*, list_*, compare_*). Descriptions are present and moderately detailed (averaging ~120 chars, within production range). However, output schemas are NOT documented, the tool descriptions state what they 'return' in natural language but provide no structured output schema definition. This is a critical gap for LLM composition: agents cannot predict response structure to plan downstream calls. Parameter descriptions are concrete (e.g., 'ISO 2 文字の国コード'), but several lack explicit format/constraint documentation. Error handling is minimal, no recovery guidance is included in responses. The codebase shows thoughtful implementation (Haversine distance, caching, language selection in GBFS autodiscovery), but the MCP surface omits output schemas entirely.

• Document output schemas for all 8 tools. For 'get_systems', add: outputSchema: { type: 'object', properties: { systems: { type: 'array', items: { type: 'object', properties: { system_id: {type: 'string'}, name: {type: 'string'}, location: {type: 'string'}, country_code: {type: 'string'}, url: {type: 'string'} }, required: ['system_id', 'name', 'location', 'country_code', 'url'] } }, total: {type: 'integer'} } }
• Add error handling context to tool descriptions. E.g., 'Returns error if system_id not found in systems.csv, try get_systems() to discover valid IDs.' Include specific error messages agents should handle.
• Add explicit bounds to numeric parameters: limit (min: 1, max: 100 for get_systems; min: 1, max: 50 for find_nearest_stations). State defaults clearly.
• Split 'system_id' parameter into two variants: one tool that accepts 'system_id' (canonical, from systems.csv) and documents this clearly; offer a separate parameter 'gbfs_url' for direct URL fallback, or remove backward-compatibility overloading and require users to call get_systems() first to resolve names.
• Add pagination to 'get_station_info' and 'get_station_status': include optional 'limit' (default: 50, max: 500) and 'offset' parameters. Return { stations: [...], total: N, offset: X, limit: Y }. Document that results beyond 500 require pagination.
• Add tool annotations for idempotency: all 8 tools are READ_ONLY and idempotent. Declare inputSchema with a property annotating this: 'readOnlyHint': true.
• Document which tools depend on prior calls. E.g., 'find_nearest_stations() requires a valid system_id from get_systems().' This guides agent planning.