An MCP server for executing SQL queries and exploring metadata in Databricks workspaces using Unity Catalog
This server exhibits significant gaps across definition quality dimensions. While all 7 tools are explicitly registered with the @mcp.tool() decorator and have descriptions, most descriptions are concise but lack depth and context. Parameter schemas are present but vary in quality, several parameters lack detailed descriptions explaining constraints, expected formats, or valid values. No output schemas are documented. Error handling is minimal, returning generic error strings rather than actionable recovery guidance. The tool set lacks idempotent operation guidance and does not include multi-step composition hints. Overall, the definitions fall into the 'Fair' (C) range with major gaps in parameter detail, output documentation, and error guidance.
Get schema information for a specific table.
Execute a SQL query on Databricks. When working with wesley_farms catalog tables, consider using get_table_relationships first to understand proper join logic.
Get predefined table join relationships from wesley_farms.gold.table_relationships. Use this when querying tables in the wesley_farms catalog to understand how tables should be joined together.
List all catalogs in the Unity Catalog metastore
List all databases/schemas in the Databricks workspace
List all schemas/databases in a specific catalog.
Output schemas not documented. All tools return free-form text strings rather than structured JSON. LLMs cannot reliably parse results or plan downstream tool calls. For example, execute_sql returns 'Columns: ... Rows returned: ...' as plain text, an LLM has to parse this unpredictably formatted string to extract column names or row counts. Best practice: return structured objects with typed fields like { columns: [...], rows: [...], total_count: 10 }.
Parameter descriptions are incomplete or missing. 'warehouse_id' in execute_sql says 'Optional' but does not explain fallback behavior or error conditions. 'query' parameter in execute_sql does not specify SQL dialect, max length, or timeout. 'database', 'table', 'catalog' parameters across tools lack format constraints (case sensitivity, special characters, etc.). LLMs cannot reliably construct valid input without explicit format guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 52 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
List all tables in a specific database.
Generic, non-actionable error messages. All tools return 'Error: ...' with the exception message. Example: 'Error executing query: ...' or 'Error listing databases: ...' does not tell the LLM what to do next, should include recovery hints like 'Check warehouse_id is valid' or 'Ensure catalog name is correct; available catalogs: [...]'. Pattern: recovery-guide.
Result limits not documented or enforced. execute_sql shows '[:10]' rows but description does not state this limit. No pagination support. Baseline pattern requires explicit documentation of result caps (typically 20-50 items) and pagination tokens (limit, offset, next_cursor). Without this, large query results may blow the context window.
Descriptions are too brief to be useful for LLM tool selection. Baseline for good descriptions is 34 - 392 chars (p10 - p90 from 549 production tools). list_databases ('List all databases/schemas in the Databricks workspace' = 60 chars) is at the lower end. list_catalogs (11 words, ~40 chars) is well below baseline. These short descriptions lack context for when to use the tool vs. a similar one, e.g., why choose list_catalogs over list_databases? No clear guidance.
No idempotency or dry-run support documented. execute_sql is a write-capable tool (can CREATE TABLE, INSERT, etc.), but description does not state whether it is read-only or if there are retry/confirmation patterns. If an LLM retries a failed INSERT, it may create duplicates. Pattern: idempotent-operation and confirmation-request.
No composition guidance. If an LLM wants to 'query tables in wesley_farms', it must call list_catalogs → list_schemas → list_tables → describe_table → execute_sql. This is tedious and not documented. Best practice: provide hints like 'Use list_tables to discover available tables, then call describe_table to understand schema before querying with execute_sql.'
Warehouse ID dependency not clearly documented. Multiple tools require DATABRICKS_WAREHOUSE_ID environment variable, but this is not mentioned in tool descriptions. If the variable is missing, tools return generic errors like 'Error: DATABRICKS_WAREHOUSE_ID not set'. This is not user-actionable, LLMs cannot configure environment variables. Consider: accept warehouse_id as an optional parameter to each tool, or document clearly that it must be set before calling.
Table name and column name parsing is fragile. describe_table code (incomplete in source) and list_tables use case-insensitive heuristics to find 'table' columns ('tablename' or 'table'). This is not documented and may fail on different Databricks configurations. Consider: use the result schema manifest explicitly rather than guessing column indices.