MotherDuck is launching Guides, a context layer for AI agents that lives inside the data warehouse. Guides are markdown files with metadata that store metric definitions, join relationships, business caveats, and style conventions — queryable via SQL and versioned like any warehouse object. On the DABStep agentic analytics benchmark, Guides improved agent accuracy by 72 percentage points and reduced cost per run by ~55% compared to agents discovering context on their own. A small model (Gemini Flash) answered 99.5% of benchmark questions correctly at roughly two cents each. Guides integrate with the MotherDuck MCP server and work with Claude, Codex, Cursor, and other MCP clients. They support org-level and user-level ownership, automatic propagation of updates, and SQL-based management via functions like MD_CREATE_GUIDE().
