Dandy AI
Request access

Schema-first project planning

The contract layer for
AI-first data development.

Projects declare typed input and output schemas. Dandy AI connects them into a dependency graph, and when a schema changes, every downstream project updates automatically — so the plan and the pipeline never quietly drift apart.

Request access

01 — Contracts that hold

Change one schema. The graph updates itself.

Every project in Dandy AI declares a typed output schema and typed input tables. Connect two projects with an edge, and that schema becomes the contract between them. Edit an output field later — rename it, retype it, add a column — and Dandy AI propagates the change through every project downstream, the same way a spreadsheet recalculates a formula.

Some nodes do more than pass data through: a join node merges two typed inputs into one deterministic output; a vector_store node adds an embedding column. The graph knows the difference between a black-box step someone still has to build and a transform it can compute for you.

Customer Events table · ingest Support Tickets table · ingest ⨝ join customer_id Churn Score

Illustrative example — not a real Dandy AI product.

directorplans the graph, reviews every schema edit
↓ spawns, one project each
subagentbuilds a project, proposes schema changes
↓ same validation path as the web UI
MCP tool callcreate_project · add_edge · update_output_schema

02 — Built for how AI teams build now

Agents plan through the same door humans do.

Dandy AI exposes its write path as MCP tools, so an agent can create products, projects, schemas, and edges conversationally — and every one of those calls runs through the identical validation and propagation logic the web UI uses. There's no separate, looser code path for automated writes.

That means a multi-agent build session leaves behind the same thing a human-planned one does: a typed, versioned graph of what was built, in what order, and why — not a pile of chat transcripts no one will reread.

Request access

Dandy AI is in gated early access.

Tell us a bit about what you'd build. We review requests by hand — no self-serve signup, no account to manage in the meantime.