One MCP server fronting several SQL Server databases
Server has 4 tools with clear verb-noun naming (list_*, describe_*, run_query). All tools have descriptions (avg 120 chars, within baseline 34-392). Input schemas are present with type definitions and parameter descriptions. However, output schemas are undocumented, tools return JSON strings without declaring field structure. Error handling is basic (text error messages, no recovery guidance). Security is well-designed (read-only enforcement via regex, no secrets in params), but descriptions lack dependency hints and don't explain when to call discovery tools first. Composition is sound: tools chain naturally (list_databases → list_tables → describe_table → run_query). Missing: pagination for large result sets, structured output documentation, and actionable error messages.
Return the columns, data types, lengths and nullability for a table.
List the SQL Server targets this server can connect to (name, server, database, read-only flag).
List base tables in a database, optionally filtered to a single schema.
Run a SQL query against a named database and return up to 1000 rows as JSON. On targets flagged read-only, anything other than a plain read is refused.
Output schemas undocumented. Tools return JSON strings without declaring field structure (e.g., list_tables returns TABLE_SCHEMA, TABLE_NAME but this is not documented). LLMs cannot plan downstream calls or extract data reliably.
run_query lacks pagination. Tool caps results at 1000 rows and warns of truncation, but offers no limit/offset parameters or cursor for agents to fetch remaining data. Large result sets blow context windows.
Error messages are text-only, not actionable. 'Error against database: ...' gives no recovery guidance. Should categorize as retryable vs fatal and suggest next steps (e.g., 'Connection lost. Retry or check database availability.').
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Descriptions lack dependency hints. list_tables and describe_table don't explain they should be called after list_databases to discover available targets. Agents may guess wrong tool order.
run_query description mentions read-only enforcement but doesn't explain what queries are rejected or how to debug. Should include examples: 'INSERT/UPDATE/DELETE rejected on read-only targets. SELECT, WITH, DECLARE, SET, PRINT allowed.'