A Model Context Protocol server that exposes a ClickHouse instance to MCP clients over stdio
Server has 5 tools with basic schemas and descriptions. All tools have names starting with action verbs (list_, describe_, get_, run_) and descriptions present. However, descriptions are minimal (10-80 chars, well below the 194-char baseline), parameter descriptions are sparse or missing, and output schemas are not documented. The run_query tool lacks critical safety documentation despite accepting arbitrary SQL. Error handling is present but not recovery-focused. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.
Get detailed information about a table including its engine, row count, size, and column definitions
Get statistics about a table including row count and disk size
List all databases in the ClickHouse instance
List all tables in a specific database
Execute a SQL query against ClickHouse. In readonly mode, only SELECT and EXPLAIN statements are allowed
Descriptions are critically short (10-80 chars vs 194-char baseline). 'List all databases in the ClickHouse instance' and 'Execute a SQL query against ClickHouse...' lack context for LLM tool selection. No guidance on when to call each tool or what structure they reveal.
Output schemas are not documented. LLMs cannot plan downstream calls or extract fields. Tests show QueryResult with Columns/Rows, but this structure is not exposed in tool definitions. Agents must infer output format from behavior.
run_query lacks safety documentation and tool annotations. Description mentions 'readonly mode' but does not explain consequences of write mode, lack of transaction support, or multi-statement rejection. No destructiveHint annotation despite WRITE risk profile.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 52 | <=2025-11-25 | v2 |
Parameter descriptions are missing or minimal. 'database' and 'table' parameters lack format guidance, constraints, or examples. 'query' parameter in run_query does not document allowed statement types, size limits, or timeout behavior.
Error messages in tests show basic rejection ('rejected') but lack recovery guidance. 'Invalid status: must be low|medium|high' pattern is absent. Agents cannot self-correct from errors without extra discovery calls.