Single tool with minimal schema and generic description. Input schema present but lacks parameter-level descriptions. Tool description is adequate (95 chars) but parameter annotation is missing. No output schema documented. Error handling returns raw exception strings without recovery guidance. No tool annotations (readOnlyHint, etc.) despite read-only semantics. Naming is clear (query_azure_sql_data) but follows verb_noun pattern only partially. No pagination, no structured output specification, no field-level documentation.

• Add parameter description to schema: 'natural_language_question' should include guidance like 'A question about products or inventory (e.g., "What products are in stock?"). Will be translated to T-SQL SELECT and executed read-only.'
• Document output schema: specify that the tool returns a string containing either a synthesized answer or an error message. Consider returning structured JSON {"answer": string, "query_executed": string, "row_count": int} for better LLM reasoning.
• Enhance error handling with recovery guidance: catch specific exceptions (SQL syntax errors, connection timeouts, permission denied) and return actionable messages like 'SQL syntax error: {details}. Rephrase your question to avoid complex joins.' or 'Database connection timeout. Retry in 30 seconds.'
• Add readOnlyHint annotation to tool definition: @mcp.tool(readOnlyHint=True) to signal to agents that this tool is safe to call without side effects.
• Add input validation: document expected question length (e.g., 10-500 chars), reject obviously malicious patterns, and return clear validation errors rather than letting bad input reach the LLM.