The Sitecore MCP server demonstrates solid definition quality with mostly well-articulated tool descriptions and proper parameter documentation. All 6 tools have descriptions exceeding the 20-character minimum. Parameters are typed with clear enums where appropriate (e.g., 'action' in indexing-set-search-index-state). Tool naming follows verb_noun conventions (get_*, rebuild_*, set_*). However, output schemas are not explicitly documented in the provided source code, descriptions mention what is returned but formal schema definitions are not visible. Error handling guidance is present in descriptions (e.g., 'Run an actual query with indexing-find-item before concluding an index is broken') but lacks structured recovery patterns. The server uses tool annotations (destructive/write hints via src/tool-annotations.ts) which is current-spec compliant. A few tools have parameter dependencies that could be more explicitly documented.

• Add explicit output schema documentation for all tools. Include response field names, types, and descriptions. For example, 'indexing-get-search-index' should document the structure of returned index objects (e.g., { name: string, state: 'Started'|'Stopped', documentCount: number, isHealthy: boolean }).
• Expand common-restart-application description from 38 characters to 80-150 characters. Clarify when this tool should be called, what it does to in-flight requests, and whether it has prerequisites (e.g., must have admin permissions). Add a non-empty input schema, even if empty: {} with a description explaining parameters are not required.
• Formalize parameter dependencies in schema using oneOf/allOf/anyOf patterns or explicit constraint descriptions. For 'indexing-rebuild-search-index', explicitly state: 'Supply id or path (but not both), or neither for whole-index rebuild. Omit name to rebuild every index (expensive).'
• Implement structured error handling with recovery paths. For each tool, define what errors are possible, what they mean, and how the LLM should respond. Example: 'If you see 'index not found', use indexing-get-search-index to list available indexes, then retry with a valid name.'
• Document return types in a structured format (JSON Schema or TypeScript interfaces). At minimum, add a 'Returns' section to each description: 'Returns: { name: string, state: Started|Stopped, documents: number, healthy: boolean }'.
• Consider wrapping mutually-exclusive parameters into a discriminated union: instead of separate id/path/neither logic, accept a mode parameter (all | subtree_by_id | subtree_by_path) that gates the relevant parameter.