Single tool 'search_cards' has a well-structured schema with proper enums, nullable fields, and pagination support. Description is comprehensive (194 chars) and explains fuzzy matching, semantic search, and language handling. Input schema is complete with type definitions, constraints (limit 1-50, offset ≥0), and enum values for type/category/rarity/stage. However, output schema is not documented, the tool returns JSON via structuredContent but no schema is provided to guide LLM on response structure. Error handling returns isError flag but lacks recovery guidance. Tool naming is clear and verb-based (search_cards). Overall solid for a single-tool server, but missing output documentation and error recovery patterns.

• Document the output schema explicitly. Define the structure of search results: e.g., {cards: [{id, name, type, rarity, image, ...}], total, nextOffset}. Include field descriptions so LLM knows what data is available.
• Add recovery guidance to error responses. Instead of 'Invalid language', return 'Invalid language. Supported: en, ja, zh. Try search_cards with language="en".'
• Clarify semantic search behavior. Specify whether 'effect' and 'attack' parameters return top-N ranked results or filter to exact substring matches. Example: 'effect: "heal" returns cards with healing abilities, ranked by relevance.'
• Add enum of supported languages to the language parameter schema. Currently it accepts any string; constrain it to actual supported values (e.g., ["en", "ja", "zh"]).
• Document the 50-result limit in the tool description. State: 'Results are capped at 50 per request to protect context window. Use offset for pagination.'