GraphRAG Grounding for Agents
Stop agents guessing your graph schema - give them the governed model over MCP.
An agent that samples nodes invents a schema; a query against a hallucinated schema doesn't crash - it returns something plausible. This recipe imports your Neo4j schema so every governed fact - Types from node labels, Elements with declared property types, NotNull/Unique constraints, references from outgoing typed relationships (multi-label targets honestly flagged neo4j.polymorphic) - is queryable by agents over MCP. Agents then MERGE on real unique keys, traverse relationships in the direction that exists, and filter on governed taxonomy values. Read-only by design: import is additive, and CoreModels holds no Neo4j credentials, so there is no path from 'consulted the model' to 'touched your graph'.