MCP Directory

Databricks MCP Servers

Official

Databricks Labs server exposing Unity Catalog functions, vector search indexes and Genie spaces as agent tools.

Unverified
stdio (local)
OAuth
Stale
Python
View repo 93 Website

Add to your client

Copy the config for your MCP client and paste it into its config file.

Install / run
git clone https://github.com/databrickslabs/mcp && uv --directory ./mcp run unitycatalog-mcp -s your_catalog.your_schema

Paste into ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "databricks-mcp-servers": {
      "command": "/path/to/uv/executable/uv",
      "args": [
        "--directory",
        "/path/to/this/repo",
        "run",
        "unitycatalog-mcp",
        "-s",
        "your_catalog.your_schema",
        "-g",
        "genie_space_id_1,genie_space_id_2"
      ]
    }
  }
}

Step-by-step guides: Add to Claude Desktop · Add to Cursor · Add to Windsurf

Before you start

  • uv installed, plus Python 3.12 via `uv python install 3.12`
  • Databricks CLI credentials configured with access to the required APIs
  • A Unity Catalog schema (and optionally Genie space IDs) the principal can access
  • A clone of the databrickslabs/mcp repo — the server runs from source via uv

About Databricks MCP Servers

The Unity Catalog server takes a schema (-s catalog.schema) and optional Genie space IDs (-g id1,id2) and builds its tool list at startup from what it finds: UC functions surface with their real names, arguments and return types; each vector search index gets a query tool; Genie spaces get tools for managing conversations and asking questions. There is no hand-written tool list to maintain — governance and the toolset both live in Unity Catalog.

Local usage is stdio through uv from a clone of the repo (there's no pip one-liner in the README): install uv, uv python install 3.12, configure Databricks credentials, and point your MCP client at uv --directory /path/to/repo run unitycatalog-mcp. For shared access you can deploy the same server on Databricks Apps via databricks bundle or databricks apps, then connect over streamable HTTP at the app URL plus /api/mcp/ (trailing slash required) with an Authorization: Bearer token.

Where it's strong: turning governed enterprise data into agent tools without glue code. UC functions are already typed and permissioned, vector indexes give RAG-style retrieval, and Genie handles natural-language analytics within a space — the app's service principal needs explicit grants for all of it.

The trade-off is status: Beta, deprecated in favor of Managed MCP servers, and last pushed mid-2025. It remains useful for self-managed setups and as a reference implementation (the repo includes a custom-server example for Databricks Apps), but new projects on current workspaces should evaluate the managed servers first.

When to use it

  • Expose a schema's Unity Catalog functions to an agent as callable tools with UC governance intact
  • Let an agent query a vector search index for RAG over enterprise documents
  • Ask Genie spaces analytics questions from inside an agent conversation
  • Prototype a data agent self-managed before moving to Databricks Managed MCP servers

Quick setup

  1. 1Clone the repo: `git clone https://github.com/databrickslabs/mcp`
  2. 2Install uv and run `uv python install 3.12`
  3. 3Configure Databricks credentials (databricks CLI auth / profile)
  4. 4Add the MCP config pointing uv at the repo with `-s your_catalog.your_schema` and optional `-g` Genie space IDs
  5. 5Optional remote setup: deploy to Databricks Apps and connect to https://<app-url>/api/mcp/ with a Bearer token from `databricks auth token`

Security notes

Unity Catalog permissions are always enforced — an agent can only call functions and query indexes its principal has been granted; credentials come from your local Databricks CLI auth, not the config file. Note the UC server is deprecated in favor of Databricks Managed MCP servers, and Labs projects are provided as-is with no support SLA.

Databricks MCP Servers FAQ

Is this the official Databricks MCP server?

It's from Databricks Labs — Databricks' experimental arm — so official but explicitly unsupported (no SLA, as-is). The fully supported path today is Databricks' Managed MCP servers built into the platform; this repo's UC server is the deprecated self-managed predecessor.

Is it safe to point at production data?

Access control stays with Unity Catalog: the server cannot exceed the permissions of the principal it runs as, and Databricks states permissions are always enforced. Scope the -s schema narrowly and run the Apps deployment under a least-privilege service principal.

Do I have to deploy anything to my workspace?

No — local stdio use only needs uv, a repo clone, and your CLI credentials. The Databricks Apps deployment is optional, for sharing one server across users or connecting remote clients over HTTP.

#databricks#unity-catalog#genie#vector-search#data

Alternatives to Databricks MCP Servers

Compare all alternatives →
Memory (Knowledge Graph)
AI, Data & Knowledge
74k

Official MCP server providing persistent, file-backed knowledge-graph memory across sessions.

Verified
stdio (local)
No auth
TypeScript
9 tools
Updated 5 months agoRepo
Sequential Thinking
AI, Data & Knowledge
62k

Structured step-by-step reasoning tool for breaking problems into revisable thought sequences.

Verified
stdio (local)
No auth
TypeScript
1 tool
Updated 6 months agoRepo
OpenMemory MCP
AI, Data & Knowledge
60k

Mem0's local-first memory layer: a Dockerized MCP server plus dashboard that keeps agent memories on your machine.

Unverified
SSE (remote)
API key
Python
5 tools
Updated 1 day agoRepo

Compare Databricks MCP Servers with:

vs Memory (Knowledge Graph)vs Sequential Thinkingvs OpenMemory MCPvs Graphiti MCP