Creativity · MCP — server

MCP Databricks Server

The MCP Databricks server connects LLM clients to a Databricks workspace. Databricks has released an official server that covers Unity Catalog browsing, SQL Warehouse execution, jobs API, and Mosaic AI model endpoints. It's ideal for data-engineer assistants that need to peek at tables, run diagnostic queries, or spin up a notebook job from chat. Authenticate with a personal access token or OAuth M2M for service scenarios.

MCP facts

Kind: server
Ecosystem: anthropic-mcp
Language: Python
Transports: stdio, http, sse

Capabilities

Tools: run_sql, list_catalogs, describe_table, list_jobs, trigger_job, query_vector_search
Resources: catalogs, schemas, tables, and job runs
Auth: Databricks PAT or OAuth M2M service principal

Install

uvx databricks-mcp-server

Configuration

{
  "mcpServers": {
    "databricks": {
      "command": "uvx",
      "args": ["databricks-mcp-server"],
      "env": {
        "DATABRICKS_HOST": "https://adb-xxxx.azuredatabricks.net",
        "DATABRICKS_TOKEN": "dapi...",
        "DATABRICKS_WAREHOUSE_ID": "abc123"
      }
    }
  }
}

Frequently asked questions

What can the MCP Databricks server do?

Browse Unity Catalog, run SQL through a Warehouse, trigger and monitor jobs, and query Vector Search endpoints. Exact tool set varies by version — check the repo for the latest capabilities.

Can I use it with Databricks Apps?

Yes. Databricks Apps hosts HTTP-transport MCP servers so you can share them with a team without everyone needing a local install.

How do I authenticate for a shared server?

Use an OAuth machine-to-machine (M2M) service principal with catalog-level grants. Avoid reusing a human PAT in shared deployments.

Sources

Databricks MCP documentation — accessed 2026-04-20
Databricks MCP GitHub — accessed 2026-04-20