Server provides 4 tools with reasonable verb-starting names and largely complete schemas. All tools have descriptions and input parameters are well-documented with types. However, output schemas are not explicitly documented, error handling guidance is absent, and some parameter descriptions could be more actionable. Tool descriptions are adequate but fall short of LLM-optimized length (target 50-200 chars; most here are 150-250+ chars). The server uses Zod for schema validation which is good, but the output structure (wrapping in {content: [{type: 'text', text: JSON.stringify(...)}]}) returns unstructured JSON strings rather than typed response schemas, reducing machine-parseability.

• Add explicit output schema documentation for each tool. Example for search_wines: 'Returns array of wine objects with fields: id (string), name (string), vintage (number), producer (string), type (string), varietal (string), score (number), drinking_window_start (number), drinking_window_end (number), cost (number), bottle_count (number).'
• Implement pagination for search_wines and search_bottles: add 'limit' (1-100, default 20) and 'offset' (0-10000) parameters; return {results: [...], total_count: number, has_more: boolean}.
• Convert unstructured response format. Instead of {content: [{type: 'text', text: JSON.stringify(...)}]}, return typed objects directly. MCP supports TextContent, but for structured data, return the object itself so LLM can extract fields without string parsing.
• Replace example-based descriptions with enum constraints where applicable. E.g., 'type' parameter: change 'Filter by wine type, e.g. Red, White, Sparkling (partial match)' to an enum ['Red', 'White', 'Rosé', 'Sparkling', 'Fortified'] if the set is known, OR document it as free-form with validation rules.
• Add error handling and recovery guidance. E.g., 'If search_wines returns no results, call get_cellar_stats to verify the cellar is populated; if empty, call refresh_data to sync from CellarTracker.'
• Add a dry_run parameter (boolean, default false) to refresh_data. When true, return what would be fetched without writing to the database.
• Add tool annotations via MCP capabilities: mark get_cellar_stats, search_wines, search_bottles as readOnly; mark refresh_data as destructive and non-idempotent (each call overwrites data).