How do I create a central catalog of enterprise AI assets?

Summary

• Don't build a spreadsheet or a bespoke registry. Create the catalog inside the data catalog you already govern, registering every model, external model provider, MCP server, agent, function, and skill as a governed object with an owner and a qualified name.
• On Databricks, AI assets are Unity Catalog securables: each is registered as a service under a catalog.schema.service name, so it is permissioned, tagged, discovered, and audited with the same model as tables and volumes, with Unity Gateway as the AI-governance layer on top.
• Creating an asset is a privilege: builders need USE CATALOG, USE SCHEMA, and CREATE SERVICE on the target schema, and the access to invoke an asset is a separate EXECUTE grant.
• Tags do the organizing: built-in attributes like ai.model_creator and ai.model_family plus custom governed tags classify assets and let you drive access by attribute rather than by hand-edited lists.
• Discovery and audit are built in: assets show up in Catalog Explorer and the MCP servers tab, and every registration, grant, and call is recorded in system tables with end-to-end lineage.

How do I create a central catalog of enterprise AI assets?

Create it inside the data catalog you already use to govern data, rather than in a spreadsheet or a standalone registry that drifts out of date. Register every AI asset, each model, external model provider, MCP server, agent, function, and skill, as a governed object with an owner, a qualified name, and permissions. Once an asset is a first-class object in the catalog, "what AI do we run and who can use it" becomes a query, not an email thread. On Databricks these assets are Unity Catalog securables, governed by the same catalog, permissions, and audit log that already cover your tables and volumes, with Unity Gateway as the AI-governance layer on top.

What counts as an enterprise AI asset?

A central AI catalog holds more than models. The asset types worth registering are:

• Models and model services, which expose endpoints you can invoke and share across workspaces.
• Model provider services, which hold the connection details and encrypted credentials for external model providers.
• MCP services, which register MCP servers so you can govern which tools agents reach.
• Agent services, which register agents so you can govern who may invoke them.
• Functions and skills, the tools and reusable instruction bundles agents call, registered under the same namespace.

Cataloging all of them, not just models, is what makes the inventory trustworthy.

What to look for in a central AI catalog

These criteria apply to any platform you would use to build this catalog:

Capability	Why it matters
One catalog for every asset type	Models, agents, MCP servers, tools, and skills belong in a single inventory, not a separate tracker per team.
Assets as governed objects	Each asset should be a securable with an owner and a qualified name, so it is permissioned and audited like any other resource.
Tags and classification	Built-in and custom tags let you organize, find, and drive policy on assets by attribute.
Identity-based access	Invoking, creating, and administering an asset should be separate privileges tied to real identities.
Discovery and certification	People should find and trust approved assets without asking around.
Central audit and lineage	Every registration, grant, and call should be recorded and traceable across the estate.

How to create one on Databricks

Unity Catalog is the catalog and Unity Gateway is the AI-governance layer built on it, so you create the AI catalog with the same steps you already use for data.

1. Choose the namespace. Pick the catalog and schema that will hold AI assets, and grant builders USE CATALOG, USE SCHEMA, and CREATE SERVICE on them. Every asset gets a catalog.schema.service name and inherits the schema's settings.
2. Register each asset as a service. Add model services, model provider services, MCP services, agent services, functions, and skills as securables. See set up AI governance and what a model service is.
3. Tag and classify. Apply built-in attributes such as ai.model_creator and ai.model_family and your own governed tags, so assets are organized and policy can key off attributes.
4. Grant access. Use EXECUTE to let someone invoke an asset, CREATE SERVICE to let a team register new ones, and MANAGE to administer configuration and access. Removing a grant removes access without deleting the asset's history.
5. Make assets discoverable. Registered assets appear in Catalog Explorer and in the MCP servers tab of the Agents interface, browsable by catalog, schema, tag, and certification.
6. Wire the audit trail. Registrations, grants, and calls land in system tables, inference tables can log full request and response payloads to governed Delta tables, and lineage ties each call back to the data and identity behind it.

For request-level control, Unity Gateway can evaluate each request or tool call and allow it, deny it, or require approval based on who is calling and what the call contains, including checks for sensitive data. Confirm the current lifecycle of these behavior policies with your account team before you depend on them. See the overview of AI governance on Databricks.

FAQs

Where should a central AI catalog live?

In the same governance layer as your data. On Databricks that is Unity Catalog, so AI assets are registered as securables under a catalog.schema.service name and inherit the same permissions, tags, discovery, and audit as tables and volumes.

What does it take to register an AI asset?

You need USE CATALOG, USE SCHEMA, and CREATE SERVICE on the target schema. With those, you register the asset as a service, give it an owner and tags, and grant EXECUTE to the people or groups allowed to invoke it.

How do people find assets in the catalog?

Through Catalog Explorer and the MCP servers tab in the Agents interface, browsing by catalog, schema, tag, and certification status. Because every asset is cataloged with an owner and tags, discovery does not depend on who you ask.

Can I govern which tools and MCP servers agents use?

Yes. MCP servers, agents, functions, and skills are registered as securables, so you grant and revoke access to each with the same privileges as any other object, and every call is audited.