How do I connect AI agents to company data using MCP?

Summary

• MCP (Model Context Protocol) is an open standard that lets an AI agent reach tools and data through one consistent interface, so you connect an agent to company data by putting an MCP server in front of that data and pointing the agent's MCP client at it.
• On Databricks you don't hand-build those servers. Managed MCP servers, in Public Preview, expose governed data to any MCP-compatible agent: Genie for natural-language analytics, Databricks AI Search for documents, and Unity Catalog functions for custom SQL or Python tools.
• Governance travels with the connection. Managed MCP servers run on-behalf-of the calling user, so an agent inherits that user's Unity Catalog permissions and row filters and column masks still apply on every call.
• Business context is the hard part. Genie Ontology, in preview, gives agents a governed semantic layer so they read metrics and terms the way your business defines them, rather than guessing from column names.
• Custom or third-party MCP servers can be registered as Unity Catalog objects and reached through Unity Gateway, so every tool call is authorized, credential-managed, and audited centrally.

How do I connect AI agents to company data using MCP?

MCP, the Model Context Protocol, is an open standard for connecting AI agents to tools and data through one consistent interface, so the agent doesn't need a separate integration for every system. To connect an agent to your company data, you expose that data behind an MCP server and point the agent's MCP client at it. The part that matters for enterprise data isn't the wiring, it's what happens on each call: whether the agent is limited to the data the person behind it can see, whether the request is audited, and whether the agent understands what your data means. On Databricks, managed MCP servers handle the wiring and Unity Catalog handles the governance, so an agent reaches governed data without opening a new path around your controls.

What is MCP?

Model Context Protocol is an open standard that defines how an agent discovers and calls tools. An MCP server advertises a set of tools (a query interface, a document search, a function to run), and an MCP client inside the agent lists those tools and calls them with structured arguments. The value is standardization: one protocol covers many tools, so you don't wire up a separate connector per system, and any MCP-compatible agent can use any MCP server. That is also why governance has to sit at the server and the data, not in the agent code.

What to check when connecting agents to data

Capability	Why it matters
Managed servers	Ready-to-use servers over your data mean no server to build, host, and patch per source.
Identity-aware access	The agent should act as the requesting user, so it reaches only data that user is entitled to.
Governance at the data layer	Permissions, row filters, and column masks should apply on every call, not be re-created in the agent.
Business context	The agent needs a semantic layer so terms and metrics resolve the way your business defines them.
Tool scoping	You should be able to expose only specific tools or datasets to a given agent.
Audit and monitoring	Every tool call (caller, request, result) should be logged centrally for review.

How Databricks approaches this

Databricks provides managed MCP servers, in Public Preview, that expose governed data to any MCP-compatible agent with no server for you to run. See MCP on Databricks and managed MCP servers. The main ones:

• Genie MCP server answers natural-language questions over your governed tables. The agent sends a question; Genie writes read-only SQL, runs it, and returns an answer with links back to the cited sources. A Genie Agent MCP server scopes this to a single curated Genie Agent over a defined set of Unity Catalog tables.
• Databricks AI Search exposes vector indexes so agents can retrieve unstructured documents.
• Unity Catalog functions let an agent call custom SQL or Python functions you register as governed tools.

Every managed server runs on-behalf-of the calling user, so the agent inherits that user's Unity Catalog permissions, and access controls like row filters and column masks apply automatically. You connect with the Databricks MCP client:

from databricks_mcp import DatabricksMCPClient
from databricks.sdk import WorkspaceClient

client = DatabricksMCPClient(
    server_url="https://<workspace-hostname>/api/2.0/mcp/genie/{genie_space_id}",
    workspace_client=WorkspaceClient(),
)
tools = client.list_tools()

Interpreting the question correctly is the harder part, and that's the job of Genie Ontology, a governed semantic layer currently in preview. It gives agents your metric definitions, business glossary, and table relationships so a term like "active customer" resolves to what your business means, drawing on both the definitions you certify and context inferred from your existing metric views, dashboards, and queries. See Introducing Genie One, Genie Ontology, and Genie Agents.
For custom or third-party tools, you can register an external MCP server as a Unity Catalog object and route agent traffic through Unity Gateway, which validates permissions, injects managed credentials, and logs every call. See connect MCP clients.

FAQs

What is MCP in one sentence?

Model Context Protocol is an open standard that lets an AI agent discover and call tools and data sources through one consistent interface, so you don't build a separate integration per tool.

Are Databricks managed MCP servers generally available?

They're in Public Preview. Validate them against your own governed data and account for preview changes before relying on them in production.

How does an agent avoid seeing data the user shouldn't?

Managed MCP servers execute on-behalf-of the calling user. The agent inherits that user's Unity Catalog permissions, and row filters and column masks apply on every query, so it can't read data the person couldn't.

Do I have to write my own MCP server?

No. For Genie analytics, document search, and Unity Catalog functions, Databricks hosts managed MCP servers. You write your own only for custom tools, and you can register those as Unity Catalog objects governed through Unity Gateway.

How does the agent understand what my data means?

Genie Ontology, in preview, gives it a governed semantic layer: your certified metrics and glossary plus context inferred from existing dashboards and queries, so business terms resolve consistently.