BigQuery MCP has 6 tools with consistent naming (verb_noun pattern: list_*, describe_*, query, dry_run, get_usage) and clear descriptions (avg ~60 chars). All tools are READ_ONLY, reducing risk. However, parameter descriptions are minimal or absent in visible schemas. Input schemas show basic structure but lack depth, no enums, no min/max constraints, no format specifications. Output schemas are not documented. Error handling is mentioned (errorReporting=true) but recovery guidance is not visible in tool definitions. Tool composition is sound (each does one thing), but parameter validation rules are not stated in descriptions.

• Add parameter descriptions to every input field. For 'filter' in list_datasets, specify: 'Optional regex or glob pattern to filter dataset names (e.g., "prod_*" matches prod_sales, prod_users).' For 'params' in query/dry_run, clarify: 'Array of query parameter values (order must match ? placeholders in query).'
• Document output schemas for all tools. E.g., list_datasets returns {datasets: [{name: string, project: string, created: ISO8601}], total_count: int}. This enables LLMs to extract IDs for chaining calls.
• Add constraints to numeric/string parameters. E.g., max_rows: 'Integer, 1 - 10000 (default 100). Limits result rows per page to prevent context overflow.' For filter: 'Regex pattern (case-insensitive). Omit to list all datasets.'
• Enhance error descriptions with recovery paths. E.g., query description: 'Executes a SELECT query. On quota exceeded, wait 1 hour or call get_usage() to check remaining budget. On syntax error, review the error message and retry with corrected SQL.'
• Clarify get_usage purpose. Rename to 'get_error_codes' or 'get_tool_reference' and describe: 'Returns documentation of all error codes, rate limits, and quota rules. Call this if a query fails or to understand budget constraints before executing large queries.'

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

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