12 tools across 3 files with mixed quality. All tools have descriptions (good) and input schemas (good), but descriptions are shallow and lack LLM-optimized guidance. Parameter descriptions are minimal. No output schemas documented. No error handling guidance. Security concern: rag_search exposes OpenAI API key and PostgreSQL connection string as parameters (critical pattern violation). Tools follow verb_noun naming adequately. Baseline: production tools average 194 chars for descriptions; these are 40-120 chars. 100% of A+ tools document return types; this server documents none.

• Move openai_api_key and conn_string from rag_search parameters to environment variables or config injection. Use os.environ.get() server-side. Update description: 'Perform RAG search using pre-configured PostgreSQL+PGVector+OpenAI. No credentials required.'
• Document output schema for each tool. E.g., list_buckets returns {buckets: [{name: string, created_date: ISO8601, size_bytes: number}], total_count: number}. Include in docstring or separate schema file.
• Expand descriptions to 150-250 chars. Add context: 'Use list_schemas first to discover available schemas, then call list_tables to see tables within a schema. Call describe_table to get column details before executing queries.'
• Add descriptions to all parameters. For query.parameters, specify: 'Array of values to substitute into parameterized SQL (e.g., [\"value1\", 123]). Used with $1, $2 placeholders in sql.'
• For query tool (WRITE risk), add confirmation step or dry-run pattern. Return analysis of what would change before executing. Document: 'Executes arbitrary SQL. Non-SELECT queries are logged for audit. Recommend preview_sql tool first.'
• Document pagination for list_objects: 'Set max_keys to 100 for first page. If result.is_truncated=true, pass result.next_continuation_token as continuation_token param.' Return continuation_token in response.
• For search_objects, clarify that max_results defaults to 50 and is capped at 1000. Add: 'Results ranked by prefix match first, then filename substring match.'
• Add resource chaining fields. E.g., list_objects response should include bucket name returned; search_objects should return bucket and key together so agent can immediately pass to get_object_content.

Inferred effective spec: 2025-06-18+.

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