MCPg demonstrates strong definition quality with 8 well-named tools, comprehensive descriptions, and explicit JSON schemas for all parameters. All tools follow verb_noun naming conventions (get_, describe_, list_, run_). Descriptions are detailed and contextual (average ~180 chars), exceeding the baseline. All parameters include type definitions and descriptions. However, output schemas are not explicitly documented in the source code provided, and some tools lack error handling guidance. The tool suite is well-composed with clear single responsibilities and strong parameter documentation including optional database targeting.

• Document explicit output schemas for all 8 tools. For example, run_select should document that it returns {rows: [{<column>: <value>, ...}], row_count: int, columns: [string]}. This enables LLMs to extract and chain results correctly.
• Add actionable error messages to run_update. Replace generic failures with guidance like: 'SQL syntax error at line 2. Check your UPDATE syntax matches PostgreSQL 14+ syntax. Retryable: no.' This helps LLMs self-correct.
• Add a dry_run parameter to run_update (boolean, default false). When true, return the query plan and affected row count without committing changes. This prevents accidental data modification.
• Enhance run_select and run_update descriptions with explicit result limiting behavior: 'Results are capped at 10,000 rows. For larger datasets, use pagination via LIMIT/OFFSET in your SQL query.'
• Document SQL injection protections in run_select and run_update descriptions: 'Queries are executed with parameterized prepared statements. Never concatenate user input directly into SQL; use bound parameters instead.' This sets LLM expectations.
• Add an idempotentHint or destructiveHint annotation to run_update to signal to MCP clients that this operation is state-modifying and may have irreversible consequences.
• Clarify the behavior of the 'fresh' parameter in run_select: does cache bypass apply to all databases, or only the primary? What is the cache TTL? This helps agents decide when to set fresh=true.
• For describe_tool and describe_self, document whether they return a paginated list or a complete catalog. If describe_self returns the full schema of all tools, cap it to avoid context window issues.