From ingestion to execution in five steps

The landing page compresses this into three steps — connect your models, map it together with AI, point your agent at it. Here are the five the platform actually runs, and what each one does with the material you already have.

1. 1

   Connect and ingest

   CoreModels reads the structure you already publish. A dbt project writes a manifest; a warehouse exposes its information schema; an API is described by a JSON Schema or a Protocol Buffers file; a content team maintains a schema.org vocabulary or a SKOS taxonomy; a research study lives in a REDCap codebook. Each connector knows the artifact its source already produces and imports it as-is — no passwords, no migration, no rebuild. Connect one, or connect them all.

   Every connector carries recipes: named, pre-scoped entry points that solve one recognizable problem for one kind of team, with a walkthrough and no requirement to learn a new vocabulary first. See the connector catalog · Read the guides

2. 2

   Model and map

   Imported structure becomes a core model: types, elements, taxonomies and the relationships between them, in a graph you can see. Then the mapping. CoreModels proposes the joins, the equivalences and the relationships it can infer across every source — the same customer under three names, the same status vocabulary with two spellings, the field that corresponds to a public standard. Compare versions side by side. Reconcile mismatches.

   A transformation engine sits underneath: one neutral representation between schema formats, so a model imported from one can be projected into another with an honest account of what the target format cannot hold. The Schema Converter

3. 3

   Collaborate and govern

   Your human team talks the proposals over and commits, in a shared workspace, with AI as collaborator. Not every field needs a ruling; the calls that carry consequence get one. Validation rules — required attributes, custom relationships, constraints, permitted values — catch errors before they propagate. Every decision is versioned and traced, so a definition has an author, a date and a reason.

   This is human-in-the-model: people set the meaning once, here, and are nowhere near the request path afterwards.

4. 4

   Expose and execute

   The governed model is served to everything that needs it. Agents — Claude, Cursor, your own — read it through the MCP endpoint on every plan, using the same permissions as the person whose credentials they run under, and can write back within those permissions; every write is versioned and reviewable like anyone else’s. Pipelines and applications read the same model through the API. Contracts, tests and downstream artifacts are generated from the governed definitions rather than typed by hand.

   Each agent gets up to 15 concurrent in-flight requests; Enterprise plans set throughput in contract. The MCP server and tool reference in the docs

5. 5

   Adapt and scale

   Models change because the business changes. Versioned schemas, branching workflows and automated validation keep the model consistent as it evolves; when a source renames a tier on Tuesday, the mapping is republished and everything reading the model changes with it. Extend the meta-model templates, track the standards your model conforms to as they move, and add new formats as they arrive.

Collaboration without bottlenecks

What an agent sees