Atlas source catalogue¶

graph_catalog is the canonical read surface for Atlas's provider picker and connection-profile status. The same operation is available through MCP and the Graph-OS REST route POST /graph/catalog; the REST adapter dispatches through the same graph_catalog implementation as MCP.

The default request (action: "list") adds a sources projection to the existing modality catalogue. Each source descriptor has availability, reason, queryMode, dialects, sync, and capabilities. Connection names and profile references are metadata only. Endpoint/DSN values, users, credentials, and probe exception text never cross this boundary.

The initial governed provider set is:

PostgreSQL and generic relational database (queryMode: sql)
registered external graph kinds (Neo4j, Apache AGE, Ladybug, Epistemic Graph, GraphQL; queryMode: cypher/graphql), with generic OpenCypher and PuppyGraph remaining unverified until an explicit read probe
virtual graphs (queryMode: uql), unverified until probed
Spark (queryMode: compute, compute-only)
Iceberg and Trino (queryMode: sql)
S3 and generic object stores (queryMode: object_store)
Teradata (availability: unsupported until a certified connector exists)

action: "preview_sync" accepts one source, one mode (delta, full, or reconcile), bounded ids, and optional neutral connection/graph selectors. It returns atlas-source-sync-preview.v1 with wouldExecute: false; execution remains exclusively through the existing source_sync tool.

The querySurfaces projection points Atlas at the existing graph_ask/nl_query natural-language paths and the engine_query(action="uql") path. Provider catalogue discovery reuses graph_configure's process-owned connection registry and the executable source-connector registry; it does not maintain a WebUI-local provider list. Graph OS owns the typed connector/source control-plane contracts, which are not a live process registry here, so no unbacked control-plane state is claimed; their runtime adapter remains a follow-up integration seam.