The ThoughtSpot MCP server shows mixed definition quality. While most tools have descriptions and some parameter documentation, there are significant gaps in schema completeness, parameter descriptions, and output documentation. Several tools lack visible input schemas entirely in the source code provided, and descriptions vary significantly in quality. Only 4 of 12 tools show complete schemas with all parameters typed and described. Error handling guidance is sparse. This is a community-grade server with noticeable gaps in production-readiness patterns.

• Add complete input AND output schemas for all 12 tools. For each, document: parameter types, constraints (enums, min/max), descriptions, and response structure (fields, types, optional/required). Example: get_answer should document response contains {answer: string, confidence?: float, source_datasource_id: string}.
• Eliminate redundant tools: merge ping() + check_connectivity() into one tool. Clarify relationship between create_liveboard() and create_dashboard(), are they identical? If so, keep one; if different, document the distinction explicitly.
• Standardize parameter naming: use snake_case exclusively. Rename 'additionalContext' → 'additional_context', 'datasourceIds' → 'datasource_ids'.
• Add recovery guidance to error scenarios. Example: 'If get_relevant_questions returns no results, try: (1) simplifying the query, (2) calling get_data_source_suggestions to verify datasource exists, (3) using send_session_message for interactive refinement.'
• Document output schemas in tool descriptions. Example: 'Returns array of question objects with fields: {id, text, suggested_datasource_id, confidence_score}.' This enables LLMs to plan multi-step flows.
• Add result limits and pagination hints to tools that return lists. search_objects accepts limit/cursor, document default (10), max (100), and guidance on when pagination is needed.
• Mark destructive operations with risk annotations: create_analysis_session, create_liveboard, create_dashboard are WRITE operations, flag them as such so agents understand side effects and retry semantics.