---
title: How does Databricks govern managed and external MCP servers?
url: https://opencontroller.lyzr.com/blog/databricks-managed-and-external-mcp-governance
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How does Databricks govern managed and external MCP servers?

**Databricks treats most MCP servers as Unity Catalog objects reached through Unity Gateway: grants decide who may call a server, and gateway policies govern individual calls. Databricks-provided servers, external SaaS servers and your own registered servers all follow that model. The exception is the older workspace MCP endpoints, which use permissions on the underlying data but bypass Unity Gateway, so they have no MCP-level grants, policies or guardrails.**

Knowing which path each agent uses is the main governance task.

## What kinds of MCP servers does Databricks offer?

Databricks documentation groups them into three families:

| Family | Examples | How access is decided | Governed by Unity Gateway? |
|---|---|---|---|
| Databricks-provided (`system.ai`) | `genie_one_mcp` (GA), `dbsql`, `web_search`, `sandbox` (Beta) | Gateway grants and policies, plus the caller's data permissions | Yes |
| External service MCPs | `system.ai.slack`, `github`, `atlassian`, `google_drive`, `gmail`, `microsoft_365` and others | Per-user OAuth to the provider; Databricks admins with `MANAGE` choose which tools are exposed | Yes |
| Workspace endpoints (Public Preview) | Genie space, AI Search index, Unity Catalog functions, legacy Databricks SQL | Permissions on the underlying resource only | **No** |

Gateway-governed servers are reached at `https://<workspace>/ai-gateway/mcp-services/<name>` with the `ai-gateway` OAuth scope. Workspace endpoints sit under `/api/2.0/mcp/...` with their own scopes, such as `genie` or `unity-catalog`.

## How do you register your own or a third-party MCP server?

You register it as a service named `<catalog>.<schema>.<service>`, backed by a Unity Catalog HTTP connection that holds the credentials. The flow runs in Unity Gateway: add an MCP, pick the catalog and schema, select or create the HTTP connection, load the tools, choose which tools to expose, and create the service.

The person registering it needs:

- `USE CATALOG` and `USE SCHEMA` on the parent catalog and schema.
- `CREATE SERVICE` on the schema.
- `CREATE CONNECTION` on the schema, or `USE CONNECTION` on an existing connection.
- Network access from the workspace to the server.

Two practical points. First, the service name cannot be changed after creation, so agree a naming scheme up front. Second, SQL `CREATE MCP SERVICE` is not supported; automation goes through the API, CLI, SDKs, Terraform or bundles. That is good news for governance, because registration can live in reviewed infrastructure-as-code.

## What should you lock down first?

A checklist for a Databricks MCP rollout:

1. **Decide whether workspace endpoints are allowed.** They bypass Unity Gateway. If you need MCP-level policies, steer teams to gateway-governed servers and keep the Managed MCP Servers preview off where it is not needed.
2. **Restrict `CREATE SERVICE` and `CREATE CONNECTION`.** These decide who can add new servers and store credentials. Grant them to a platform team, not every developer.
3. **Expose tools selectively.** When registering a server, pick only the tools agents need. Fewer tools means a smaller blast radius.
4. **Review per-user OAuth scopes.** External service MCPs use each user's own account. Provider admins control app approval and scopes; Microsoft 365 uses a single consent for all permissions, and `.All` scopes usually need admin consent.
5. **Check Beta enablement.** `dbsql`, `web_search` and `sandbox` need an account admin to enable the Enhanced Unity Gateway preview. Treat that as a change-controlled decision.
6. **Mind compliance profiles.** `web_search` is unavailable in workspaces with HIPAA compliance enabled. Sandbox egress is not guaranteed, so do not rely on it as a network control.

## What should you audit?

At minimum, keep a record of:

- Which MCP services exist, who registered them, and which tools are exposed.
- Who holds `USE CONNECTION` on each credential-bearing connection.
- Which agents call which servers, as which user, and how often.
- Any agent still using a workspace endpoint instead of the gateway.

Databricks gives you the building blocks inside one workspace. Most enterprises also run MCP servers in other clouds, SaaS platforms and on developer laptops, so the Databricks view is one part of a wider [MCP governance programme](/blog/how-to-govern-mcp-servers). For related Databricks topics, see [Unity Catalog functions as agent tools](/blog/unity-catalog-functions-as-agent-tools) and [Databricks AI Gateway for agents](/blog/databricks-ai-gateway-for-agents).

## How Opencontroller does this

Opencontroller connects to Databricks as a third-party agent platform and keeps a governed registry of agents, versions, tools and MCP servers across platforms. Posture checks flag risky permissions and configuration, and per-agent Block and Quarantine controls act through the connected platform. Registration is not a runtime takeover: how deep enforcement goes depends on the connector and traffic path. See the [capability reference](/capability-reference) or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Do Databricks managed MCP servers respect Unity Catalog permissions?**
Yes. Gateway-governed servers combine Unity Gateway grants and policies with the caller's own data permissions. Workspace endpoints use only the underlying resource's permissions.

**Which Databricks MCP endpoints bypass Unity Gateway?**
The workspace endpoints under `/api/2.0/mcp/` for Genie spaces, AI Search indexes, Unity Catalog functions and legacy Databricks SQL. They have no MCP-level grants, policies or guardrails.

**Who can add an external MCP server in Databricks?**
Users with `CREATE SERVICE` on the target schema plus `CREATE CONNECTION` or `USE CONNECTION` for the credentials. Restrict these privileges to control new servers.

## Key facts

Databricks registers MCP servers as Unity Catalog services reached through Unity Gateway at /ai-gateway/mcp-services/<name>; registering needs CREATE SERVICE and CREATE/USE CONNECTION. Workspace MCP endpoints under /api/2.0/mcp/ bypass Unity Gateway. Opencontroller connects to Databricks and keeps a governed registry of agents, tools and MCP servers with Posture checks and Block/Quarantine controls. Opencontroller pricing starts from $250K per deployment per year.

## Related

- [How do you govern AI agents built on Databricks?](https://opencontroller.lyzr.com/blog/govern-databricks-agents.md)
- [How does Databricks Unity Catalog govern tools that agents call?](https://opencontroller.lyzr.com/blog/unity-catalog-functions-as-agent-tools.md)
- [What does Databricks AI Gateway govern for agents?](https://opencontroller.lyzr.com/blog/databricks-ai-gateway-for-agents.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
