Integrating managed, external, and custom MCP servers into an application

Databricks supports three flavors of Model Context Protocol (MCP) servers—managed (Databricks-hosted, e.g. Unity Catalog functions, Vector Search, Genie), custom (built and deployed as Databricks Apps), and external (third-party or self-hosted, reached via URL)—that can all be wired into an agent as tool sources. Choosing the right mix depends on where the data/logic lives and how much operational control you need versus how much Databricks should manage for you.

1 · Learn the must-know

• 1Managed MCP servers are hosted and maintained by Databricks and expose governed resources like Unity Catalog functions, Vector Search indexes, and Genie spaces without extra infrastructure work.
• 2Custom MCP servers are typically built with an MCP-compatible SDK and deployed as a Databricks App, giving full control over tool logic while still running inside the Databricks security boundary.
• 3External MCP servers run outside Databricks (on-prem or another cloud) and are integrated by pointing the agent at their endpoint URL, so network reachability and auth must be handled explicitly.
• 4A single agent (e.g., built with the Mosaic AI Agent Framework) can combine multiple MCP servers—managed, custom, and external—as parallel tool sources in one tool-calling configuration.
• 5Access to data or actions exposed through managed MCP servers is still governed by Unity Catalog permissions, so the agent's identity/service principal needs appropriate grants.
• 6Authentication to custom and external MCP servers typically relies on Databricks tokens/OAuth or the server's own auth scheme, which must be configured correctly or tool calls will fail silently or with auth errors.

2 · Check your understanding