DatosVivos MCP server has 7 tools with READ_ONLY risk classification. Tool names follow verb_noun convention (search_, analyze_, execute_, get_, classify_, generate_, resolve_) which is positive. However, critical gaps exist: descriptions are present but lack LLM-optimization guidance (e.g., 'analyze_question' description is 180 chars but omits when/why to use it vs search_datasets); parameter descriptions are minimal or missing context about constraints; output schemas are not documented in the source code provided; no error handling guidance visible; no pagination or result-limiting patterns for search_datasets or analyze_question which could return large result sets. The 'analyze_question' tool appears to combine multiple responsibilities (intent classification, dataset retrieval, query generation, execution, explanation) which violates single-responsibility principle. Parameter 'dataset_id' is optional in analyze_question but no guidance on what happens if omitted. No evidence of input validation, rate limiting, or recovery guidance.

• Document return types for all tools. Specify field names, types, and whether results are paginated. Example: 'Returns array of {dataset_id: string, title: string, description: string, entity: string, record_count: integer}. Results limited to k items; use offset for pagination.'
• Add constraint descriptions to parameters. For 'query': 'Natural language search query in Spanish or English (max 500 chars)'. For 'k': 'Maximum results to return (1-100, default 5)'. For 'soql': 'Valid SoQL query string; max 10000 chars'.
• Split analyze_question into separate tools: search_datasets_for_question, generate_soql_from_question, execute_and_explain_query. Each tool does one thing; agent composes them.
• Add error handling guidance to descriptions. Example for execute_soql: 'Returns error if dataset_id not found (try search_datasets first) or soql syntax invalid (error message includes position and suggestion).'
• Implement result limiting and pagination. search_datasets should cap at 20-50 results by default; add offset/cursor parameter. Document in description: 'Results limited to k items. For more, use offset parameter (0-indexed).'
• Add tool annotations (readOnlyHint, idempotentHint) to all tools since all are READ_ONLY and idempotent. Helps agents understand retry safety.

Inferred effective spec: <=2025-11-25.

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