Skip to main content
Last updated on

AI governance with Unity AI Gateway

Beta

This page covers the new Unity AI Gateway (visible in the sidebar of the UI), which is currently in Beta. Account admins can enable access to this feature in the account console Previews page. See Manage Databricks previews.

For details on the previous version of AI Gateway (not Unity AI Gateway), see AI Gateway for serving endpoints.

Unity AI Gateway is the Databricks governance solution for enterprise AI. Built on Unity Catalog, it extends governance beyond your data and AI assets to the runtime interactions between models, agents, MCP servers, and tools. Use it to:

  • Control which AI services teams can use: Register Databricks-hosted and external models, MCP services, and agents in Unity Catalog, then grant access with standard Unity Catalog privileges. Databricks has foundation model services by default, plus managed MCP services for apps like Google Drive, Jira, Slack, and GitHub.
  • Route and manage AI traffic: Route requests, set rate limits, configure fallbacks, and manage budgets across models and MCP services. Hard spend caps stop requests when a budget is reached, rather than alerting after the fact.
  • Set guardrails and access policies: Attach service policies to allow, deny, require approval for, or transform individual requests and responses. Built-in policies protect against PII exposure, prompt injection, and unsafe content.
  • Monitor usage, cost, and risk: Track who uses which services, how much they spend, and what happens during each request, with unified agent tracing across model and MCP activity.

New to AI governance on Databricks? See Get started with AI governance for an end-to-end setup path.

Control which AI services teams can use​

Unity Catalog manages AI assets as securable objects. Register them once, then grant and revoke access using the same privileges you use for tables and volumes:

To define and share model services as Unity Catalog securable objects across workspaces, see Create and manage model services.

Route and manage AI traffic​

Unity AI Gateway routes requests to your model services and MCP services from a central control plane. You can manage capacity, availability, and spend across Databricks-hosted and external providers.

    • Apply rate limits
    • Enforce consumption limits on model services and MCP services to manage capacity and cost.
    • Manage budgets
    • Monitor spend and set per-user thresholds and hard caps across Databricks-hosted and external providers.
note

Unity AI Gateway features don't incur charges during Beta.

Set guardrails and access policies​

Unity Catalog privilege grants determine whether a principal can call an AI service. Service policies govern how that interaction proceeds, based on the content of the request and response and on who is making the call. This matters most when agents act on behalf of users and reach external systems.

A service policy is a type of attribute-based access control (ABAC) policy scoped to AI services. You can allow, deny, or require human approval for an interaction, or transform request and response content. For example, you can redact personally identifiable information (PII) with a built-in policy such as system.ai.mask_pii. See Service policies for AI securables and Create and attach a service policy.

Monitor usage, cost, and risk​

You can track activity, spend, and outcomes across all Unity AI Gateway services.

    • Monitor usage
    • Track requests, token usage, and latency for model services using system tables.
    • Analyze cost
    • Attribute Databricks cost to services, target models, principals, and tags.

How it works together​

Unity AI Gateway builds on Unity Catalog governance across three layers:

  • AI assets: Unity Catalog manages models, functions, connections, and services as securable objects, governed with standard Unity Catalog privileges. Services include model services, agent services, and MCP services.
  • AI traffic: Unity AI Gateway is the central control plane for all AI services, including foundation models, tools, and agents.
  • AI service behavior: Service policies govern the content of requests and responses to AI services, based on who is calling and what is sent.

Monitoring spans all three layers: you can track usage, cost, and risk across assets, traffic, and service behavior, including who used which service and what happened during each request.

Get started​

The following topics help you set up and apply AI governance on Databricks.

Model serving endpoints (previous)​

The previous version of AI Gateway has governance features for model serving endpoints at the workspace level, including external model endpoints, Foundation Model API endpoints, and custom model endpoints.

On this page