Skip to main content

Databricks Unity AI Gateway

Register Alice once as a Unity Catalog HTTP Connection, then attach it as an External policy to a governed model or MCP service. The gateway calls Alice before the model (model_call) and after it (model_result), or around an MCP tool (tool_call, tool_result), receives ALLOW or DENY, and enforces inline. Start in Log mode, flip to Enforce when the verdicts match your intent.

Status: Generally available · Evaluates: Prompts, Responses, Tool calls · Vendor: Databricks

Model input and model output through the gateway, and MCP tool calls and results with tool message types.

Setup​

Configuration: OAuth client id and secret, app_id

  1. Create OAuth client credentials in Alice.
  2. Create a Unity Catalog HTTP Connection: host https://api.alice.io, path /v2/evaluate/databricks, auth OAuth M2M, token endpoint https://api.alice.io/v1/oauth/token.
  3. Attach an External policy to the model, choose the phases, and set the policy configuration below.
  4. Run in Log mode, then switch to Enforce. Allow 60 to 90 seconds for a new policy to take effect.

Tool calls​

MCP tool calls and their results are evaluated only once Alice has enabled WONDER_FENCE_AGENTIC_MESSAGE_TYPES on your project, and only by policies that list the Tool Call or Tool Result message type. Without it, Alice answers ALLOW to every tool event without evaluating it.

Model events fail closed: if Alice cannot evaluate one, it answers DENY. Tool events fail open: if Alice cannot evaluate one, or cannot read its content, it answers ALLOW, so an outage does not take your MCP-backed agents offline. A missing or unknown app_id is denied on both. A long tool result is screened in parts, up to about 80,000 characters.

Example​

{ "app_id": "<your-application-id>" }

// Alice replies to the gateway
{ "result": "ALLOW" }
{ "result": "DENY", "reason": "Message violates policy: PII, Prompt Injection" }