Databricks MCP Server Cursor IDE Setup 2026: Managed MCP Endpoints

Key Takeaways

Managed MCP vs. self-hosted MCP on Databricks

Databricks documents two distinct ways an MCP server ends up connected to your workspace, and mixing them up is the most common setup mistake:

This guide is about the managed path — connecting Cursor directly to Unity Catalog, AI/Vector Search, or Genie without building anything.

Prerequisites

The three managed endpoint types

Unity Catalog Functions MCP exposes SQL or Python functions registered in Unity Catalog as callable MCP tools. The endpoint pattern is:

https://<workspace-hostname>/api/2.0/mcp/functions/<catalog>/<schema>/<function_name>

Required scope: unity-catalog.

AI/Vector Search MCP exposes a Databricks Vector Search index for semantic retrieval — the AI can search unstructured or embedded content instead of only running exact-match SQL. The endpoint pattern is:

https://<workspace-hostname>/api/2.0/mcp/ai-search/<catalog>/<schema>/<index_name>

Required scope: ai-search.

Genie MCP connects to a Genie space — Databricks' natural-language-to-SQL feature grounded in a curated data model — so the AI can ask Genie questions instead of writing raw SQL against tables it doesn't fully understand. The endpoint pattern is:

https://<workspace-hostname>/api/2.0/mcp/genie

Required scope: genie. Unlike the other two, this endpoint isn't parameterized by catalog/schema/name in the path — it talks to whichever Genie space your token has access to.

Treat the path patterns above as what Databricks documents today, not a value to copy without checking. Public Preview endpoints are the ones most likely to shift before general availability — confirm the exact URL for your workspace in the Databricks UI (workspace settings, or the catalog/schema/index page for the object you're exposing) before wiring it into Cursor.

Step 1: Generate a Databricks token

In your Databricks workspace: Settings → Developer → Access tokens, generate a personal access token scoped to an account that has the permissions you want the AI to have — not a broad workspace-admin identity. If your organization uses OAuth-based service principals instead of PATs, use that token in the same place below.

Step 2: Add the endpoint to Cursor's mcp.json

Pick the endpoint type you need. A Unity Catalog Functions example:

{
  "mcpServers": {
    "databricks-uc": {
      "type": "http",
      "url": "https://your-workspace.cloud.databricks.com/api/2.0/mcp/functions/main/analytics/get_customer_summary",
      "headers": {
        "Authorization": "Bearer YOUR_DATABRICKS_TOKEN"
      }
    }
  }
}

A Genie example, since the URL doesn't need a catalog/schema/name suffix:

{
  "mcpServers": {
    "databricks-genie": {
      "type": "http",
      "url": "https://your-workspace.cloud.databricks.com/api/2.0/mcp/genie",
      "headers": {
        "Authorization": "Bearer YOUR_DATABRICKS_TOKEN"
      }
    }
  }
}

You can run more than one Databricks entry side by side — a Unity Catalog Functions connection and a Genie connection are different mcpServers keys, each with its own URL and, if you want to scope them differently, its own token.

Step 3: Restart Cursor and verify

Restart Cursor completely, then check View → Output → MCP for a successful connection log line rather than an auth error. Test with a query matched to whichever endpoint you configured:

List the Unity Catalog functions available through the databricks-uc MCP connection.

Ask Genie: what were total orders by region last quarter?

If the AI returns real function names or a real Genie answer rather than an error, the token's scope and the URL are both correct.

Custom and third-party MCP servers on Databricks

Two paths exist beyond the three managed endpoints, worth knowing even if you don't need them today:

Troubleshooting

401 or 403 on connect. The token is missing, expired, or lacks the specific scope for the endpoint type (unity-catalog, ai-search, or genie). A 403 specifically usually means the token is valid but the account behind it lacks a grant on the catalog, schema, function, index, or Genie space in the URL — check object-level permissions in Unity Catalog, not just the token itself.

404 on the /api/2.0/mcp/... path. Either managed MCP servers aren't enabled for your workspace yet (it's Public Preview — check workspace settings), or the path segment is wrong for the endpoint type. functions and ai-search both require a catalog/schema/name suffix; genie does not. Mixing those up produces a 404 that looks like a permissions problem but isn't.

Function or index shows up but calling it errors. Confirm the function or Vector Search index actually exists and is deployed in the catalog/schema you pointed at — the MCP layer surfaces whatever Unity Catalog already has; it doesn't create anything.

Genie answers are wrong or generic. Genie's answer quality depends entirely on how well the Genie space's underlying data model and sample queries are curated in Databricks. This isn't an MCP-layer problem — improve the Genie space itself in the Databricks UI, the same as you would if you were using Genie directly without MCP in the loop.

Found a tutorial with a local npx-installed Databricks MCP server. That's describing a self-hosted or community setup, not the managed endpoints in this guide. Managed MCP servers run inside Databricks — there's nothing to install locally for the three types covered above.

Frequently Asked Questions

Q: Do I need to install or run anything locally to use Databricks MCP with Cursor?
A: No, for the three managed endpoint types (Unity Catalog Functions, AI/Vector Search, Genie). They run inside your Databricks workspace and are reached over HTTP — your mcp.json entry just needs a URL and an Authorization header.

Q: What's the difference between Unity Catalog Functions MCP and Genie MCP?
A: Unity Catalog Functions exposes specific registered SQL or Python functions as individual callable tools — precise, but only covers what's explicitly registered. Genie is natural-language-to-SQL grounded in a curated data model, so the AI can ask broader business questions without you registering a function for every possible query.

Q: Is this the same as the Databricks MCP server people mention running via a local package?
A: No. Managed MCP servers run inside Databricks with no local process. A local, package-installed server is a different, self-hosted setup — check whether the tutorial you're reading is describing the managed endpoints or a custom deployment before following its steps.

Q: Can I connect more than one Databricks endpoint type to Cursor at once?
A: Yes. Unity Catalog Functions, AI/Vector Search, and Genie are separate URLs — add each as its own key under mcpServers in the same mcp.json, with its own token if you want to scope access differently per endpoint.

Q: Is this feature stable enough for production use?
A: Managed MCP servers are Public Preview as of the documentation current at the time of writing. Public Preview features can change, including endpoint paths, before general availability — check the current Databricks docs for your cloud (AWS, GCP, or Azure) before depending on the exact URL patterns in this guide for anything business-critical.

Related Guides