Announcing managed MCP servers with Unity Catalog and Mosaic AI Integration

Model Context Protocol (MCP) has taken off in recent months with enthusiasm all across the industry as a standard for equipping LLMs with tools. This is an important step forward because it gives LLMs the context they need to take actions in a more natural form. However, enterprises all over are trying to figure out how to use it without sacrificing governance, security, and discoverability. You need to keep your data secure while equipping agents with context about your business: questions around authentication and authorization, access, and control are swirling.

We’re excited to embrace MCP and combine it with the power of Unity Catalog and Mosaic AI. This gives you the best of all worlds: MCP for your agents to take action, Mosaic AI for building and evaluating agents, and Unity Catalog for governance and discovery. Now, you can equip your agents with data intelligence in a way that respects your security and governance as an organization.

With this launch, we’re handling the hard parts of MCP for you: our managed servers support on-behalf-of-user auth out of the box, respecting the governance you’ve already established in Unity Catalog. We also support OAuth as standard in Databricks Apps for easy hosting, and Playground is a secure environment to prototype quickly.

Use Databricks’ managed servers for secure data access

Our initial set of managed servers lets you securely access data in Databricks with Genie, Vector Search, and UC Functions. Built with enterprise-grade security in mind, our managed MCP servers automatically respect a user’s permissions. This means you can continue to govern all your data in Unity Catalog and leverage MCP without yet another place to manage and govern. 

Imagine you’re building a customer support agent that helps support representatives at your telcom company. You’ll be able to leverage structured and unstructured data from Unity Catalog to make your agent smarter:

1. Genie lets you access your structured data, like account and plan information
2. Vector Search lets you access your unstructured data, like support and knowledge base articles
3. UC Functions let you build deterministic functions to answer questions around billing

These servers are also managed for you, with no required upkeep or maintenance. As the MCP standard continues to evolve, we’ll update these to support the latest and greatest. This lets you focus your energy on what matters to your organization instead.

Securely host your own MCP server with Databricks Apps

Let’s keep building on our telcom support agent: we have some internal APIs that let us know about any current outages and report new ones. However, all this information is trapped in our infrastructure without a clear way to make it accessible to our agent.

This is where Databricks Apps helps bring your agent to life: with out-of-the-box OAuth support, an easy, Git-based deployment, and permissions and governance built-in, you can turn your legacy services and APIs into MCP servers in just a few minutes. Databricks Apps are built directly on our serverless infrastructure, meaning you don’t need to worry about scalability as your agent gets more use.

Try our easy-to-use MCP server template from Marketplace to get started today.

Build and deploy agents with MCP using Agent Bricks or AI Playground

Databricks then makes it easy to take advantage of MCP to build and deploy agents. In Agent Bricks, Multi-Agent Supervisor will support the MCP servers you’ve built in Databricks Apps, letting you connect your agent with your internal services with just a click of a button.

You can also use our AI Playground to prototype agents using MCP, letting you test both MCP servers and agent logic before writing any code. Simply use the “Tools” dropdown with any tools-enabled LLM to try our managed servers or your custom servers through Databricks Apps.

Using Databricks to deploy your agents, you’ll be able to easily see which tools are being used and how with MLflow and Agent Evaluation. This helps you tweak your agent logic or tool logic to fit the needs of your use case, giving you full visibility into how your agent is behaving:

Future expansion

We’re really excited about what we’re launching, and we’re just getting started. In the future, we’ll expand support for other types of Databricks resources with managed servers, like DBSQL. We’ll also plan to build out catalog support for managing, discovering, and governing MCP servers offered by other companies and services, letting you centralize right where you do today.

Get started

Databricks managed MCP servers are now in Beta. Visit our docs to learn more about how to connect to them or host your own MCP server in Databricks Apps.