Spice MCP

Catalog Discovery Summary - There is no public REST endpoint to browse the full catalog. Discovery is best achieved using Dune SQL primitives and fallback probes. - **Native SHOW statements are preferred** - they're faster than information_schema queries. See issue #10 for details. Approach - Schemas - SHOW SCHEMAS - SHOW SCHEMAS LIKE '%keyword%' - ⚠️ Avoid: information_schema.schemata (slower, causes lag) - Tables - SHOW TABLES FROM <schema> - If SHOW is blocked, probe candidate names via SELECT 1 FROM <schema>.<table> LIMIT 1 - ⚠️ Avoid: information_schema.tables (slower, causes lag) - Columns - SHOW COLUMNS FROM <schema>.<table> - Fallback: SELECT * FROM <schema>.<table> LIMIT 1, infer columns and Polars dtypes client-side - INFORMATION_SCHEMA (Deprecated) - Previously used for portability, but causes performance issues - Native SHOW statements are now used directly (faster, no lag) - Kept as fallback only if SHOW is blocked Helpers in this repo - `src/spice_mcp/service_layer/discovery_service.py` provides service wrappers around the Dune adapter: - `find_schemas(keyword)`, `list_tables(schema, limit)`, `describe_table(schema, table)` MCP Tools - dune_discover: **PRIMARY discovery tool** - unified search across Dune API and Spellbook, returns verified tables only - Automatically parses dbt configs from Spellbook models to resolve actual Dune table names - Verifies tables exist in Dune before returning (uses persistent cache) - Filters out non-existent tables - dune_describe_table: describe columns with SHOW + fallback