Local stdio MCP server exposing a read-only SQL database (Oracle, PostgreSQL, MySQL, SQL Server, SQLite) as native Claude tools
Strong foundation with clear naming, comprehensive descriptions, and proper input schemas. All 6 tools follow verb_noun convention (test_connection, run_query, list_schemas, list_tables, describe_table, get_table_sample). Descriptions are detailed (100-250 chars) and explain WHAT, WHEN, and prerequisites. Parameters have types and descriptions. Tool annotations (readOnlyHint) are present. Main gaps: output schemas are not documented in code (only inferred from implementation), error handling lacks recovery guidance, and no pagination support despite potential for large result sets.
Return full structure of a table or view: columns (name, type, nullability, default, comment), primary key, foreign keys, indexes, and table comment.
Return a small sample of rows from a table or view. The fastest way to understand what a table actually contains. Identifiers are safely quoted; the query is read-only and row-capped.
List the schemas visible to the connected account.
List tables and views, optionally filtered by schema and/or name substring.
Execute a read-only SQL query and return columns + rows. Only a single SELECT / WITH / EXPLAIN statement is allowed; anything that could modify data, run procedural code, or lock rows is rejected. Output is capped at the server's configured row ceiling and reports `truncated: true` if more rows exist.
Verify the database connection and report the dialect and server version. Use this first to confirm credentials and network reachability before running queries.
Output schemas not documented. Tool descriptions state what is returned (e.g., 'columns + rows', 'full structure') but JSON Schema for responses is absent from code. LLMs cannot plan downstream operations without knowing response field names and types.
No pagination support. list_schemas and list_tables could return hundreds of items but lack limit/offset/cursor parameters and total_count in response. Large result sets will blow context windows.
Error handling lacks recovery guidance. run_query catches ReadOnlyViolation and raises ValueError, but error messages do not guide the LLM on what to do next (e.g., 'Use SELECT instead of UPDATE'). No categorization of retryable vs. fatal errors.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
list_schemas description is minimal ('List the schemas visible to the connected account.'). Does not explain when to call it or what structure it reveals. Should guide: 'Call this first to discover available schemas before listing tables.'
No dry-run or confirmation for destructive operations. While all tools are read-only, the server lacks a pattern for future write operations. If extended to support INSERT/UPDATE/DELETE, should implement confirmation-request pattern.