dbmcp is a STDIO-only database introspection tool with 4 read-only tools. Tool definitions are visible in dbmcp/tools.go with explicit input/output structs and JSON schema tags. However, the server has significant quality gaps: (1) Descriptions are present but generic and lack LLM-optimization guidance; (2) Parameter descriptions exist but are minimal (e.g., 'optional; database connection id' lacks context on when to use it); (3) No output schema documentation, only inferred from code; (4) No error handling guidance for the LLM (e.g., what to do if a query fails or table doesn't exist); (5) No tool annotations (readOnlyHint, idempotentHint) despite all tools being read-only; (6) No pagination support despite run_select_query potentially returning large result sets. The tools themselves are well-structured and follow a consistent pattern, but lack the depth needed for robust LLM reasoning.

• Add explicit output schema documentation to each tool. Define the structure returned by get_tables (e.g., 'Returns {tables: [string], table_count: int}'), get_database_info (e.g., 'Returns {database_name, database_vendor, database_version, connection_status}'), etc. This allows LLMs to plan what data they'll receive.
• Expand parameter descriptions with context: 'connection_id (optional, string): The ID of a configured database connection. If omitted, uses the default configured connection. Use this if the agent has access to multiple databases and needs to query a specific one.'
• Add pagination to run_select_query: introduce 'limit' (default 100, max 1000) and 'offset' parameters, and return {rows: [], row_count: int, total_available: int, has_more: bool}. Cap results at 100 rows by default to prevent context explosion.
• Enhance tool descriptions with selection context. E.g., 'get_database_info: Call first to verify the database is accessible and identify the vendor (MySQL/PostgreSQL/SQLite). Use before calling get_tables to ensure the connection is active.' This guides the LLM on sequencing.
• Add actionable error messages. When run_select_query rejects a non-SELECT query, return: 'Only SELECT queries are allowed. This tool is read-only. If you need to modify data, use a different database tool.' When describe_table fails (table not found), return: 'Table "table_name" not found. Call get_tables() to see available tables.'
• Add tool annotations: Mark all four tools with 'readOnlyHint: true' to signal to the LLM that these tools do not modify state and are safe to call without confirmation.