Hey all 👋 We just shipped Guides, a context layer for AI agents working with your MotherDuck data.
A Guide is a versioned markdown file that lives in MotherDuck: metric definitions, join conventions, and any other free-form context you'd give to an analytics agent. Your agent loads the relevant Guides through the MotherDuck MCP before it writes SQL, so it stops guessing at your schema.
On the DABStep benchmark this took agent accuracy up 72 points and cut cost per run ~55%. A small, fast model (Gemini 3 Flash) got 417 of 419 questions right with Guides in hand.
Manage them with the MCP (Claude, Codex, Cursor, etc.) or SQL (MD_CREATE_GUIDE()). With the SQL table functions, you can even source control Guides in git, which enables a nice CI workflow.
Check out the launch blog, and this docs resource for bootstrapping Guides using existing context.
We'll be landing many improvements over the next couple of weeks, including an in-UI editing experience. Would love to hear what context you put in your first Guide.
.png){kind=logo}
