Databricks - Forge

Forge connects Databricks account and workspace APIs, system tables, and Unity AI Gateway control points. Inventory, delayed telemetry, and inline Forge policy are kept as separate capability states so an endpoint is never labeled protected based on configuration or usage tables alone.

Sources

Source	Data	Control boundary
Serving Endpoints API	Endpoint, served entity, model, route, and API-type metadata	Inventory only
Gateway usage tables	Requester, model, tokens, latency, routing, and request tags	Delayed telemetry
Inference tables	Optional request and response payload logging	Observed only
Custom guardrail evaluator	Forge policy request, decision, and outcome	Inline Forge PDP
Unity Catalog MCP	Servers, connections, functions, tools, and schemas	Inventory until invocation proof
Unity Catalog tags	Table and column label taxonomy and exact data context	Policy data context

Connection

Field	Description
Account ID	Databricks account used to discover accessible workspaces
Workspace mode	Account discovery or explicit workspace-only mode
Workspace URLs/IDs	Fallback or explicitly scoped workspaces
SQL warehouse ID	Warehouse used to query supported system and information schemas
Credential	Account/workspace credential stored as a Forge-managed secret

Open Integrations → Databricks Unity AI Gateway and save the account, workspace discovery mode, credential, and SQL warehouse.
Run a sync and compare serving endpoints, Gateway configuration, system-table evidence, and MCP objects with the selected workspaces.
For inline policy, configure a Unity AI Gateway custom guardrail to call the Forge OpenAI-compatible evaluator, then run the generated evaluator canary.

Permissions

Capability	Databricks access
Endpoints	Workspace assignment and serving-endpoint visibility
Usage	SQL warehouse plus `USE`/`BROWSE`/`SELECT` on system tables
Inference tables	`SELECT` on the configured table and backing storage setup
MCP	Connection read plus `USE`/`EXECUTE` on exposed UC functions
Data labels	Visibility into selected `information_schema` tag views

Enforcement

Gateway usage and inference tables are evidence after execution. The custom guardrail evaluator is the agentless inline path: Unity AI Gateway sends the request or response to Forge, Forge evaluates the applicable policy, and the gateway receives the resulting allow, block, redact, or audit outcome.

Boundaries

MCP server or function inventory does not prove a tool invocation. Inference tables can be delayed, truncated, or metadata-only. Forge reports those states explicitly and requires an evaluator or proxy receipt before claiming policy enforcement.