Databricks MCP

What does a Databricks MCP server do?

A Databricks MCP server lets an AI assistant work with data in your workspace. Genie One takes a plain-language question, searches your data, writes SQL and returns an answer with links to its sources. Databricks SQL runs the exact SQL an agent sends. Other Databricks-provided MCPs cover web search, sandboxed code and outside services.

• Genie Agent: a narrower server for one Genie Agent, which holds up to 50 Unity Catalog tables, views or metric views. It's in Public Preview and only reads.
• AI Search and Unity Catalog functions: workspace endpoints in Public Preview.
• External services: Slack, GitHub, Jira and Confluence, Google Drive, Google Calendar, Gmail and Microsoft 365 through managed OAuth.
• Your own servers: you can add a third-party server or build one from Databricks Apps templates.

Is there an official Databricks MCP server?

Yes. Databricks provides its own MCPs in the system.ai schema, so you don't host or register a server. Claude's directory lists the Genie connector as made by Databricks. The older databrickslabs/mcp repo says its server is deprecated and recommends managed servers, and Labs projects have no formal support SLAs.

• Generally available: Genie One MCP. Its Beta endpoint, /api/2.0/mcp/genie, reaches end of life on 30 November 2026.
• Unity Gateway beta: system.ai.dbsql, system.ai.web_search and system.ai.sandbox. An account admin enables the beta on the account console Previews page.
• Public Preview: Genie Agent, AI Search, legacy Databricks SQL and Unity Catalog functions endpoints. A workspace admin enables Managed MCP Servers.

What is the Databricks MCP server URL, and how do you sign in?

There's no single URL. Each MCP sits on your workspace at https://<workspace-hostname>/ai-gateway/mcp-services/<mcp-name>, with your workspace ID in the o parameter. Genie One's name is system.ai.genie_one_mcp. You sign in with OAuth as yourself, or with a service principal or personal access token. Copy URLs from Unity Gateway > MCPs.

• Unity Catalog grants: each system.ai MCP is a Unity Catalog securable. Callers need EXECUTE on the MCP, plus USE CATALOG and USE SCHEMA on its parents.
• Unity Catalog objects: a Genie Agent is a set of up to 50 Unity Catalog objects, and AI Search and function endpoints name a catalog and schema in their URL.
• Databricks MCP endpoints don't support dynamic client registration, so the client needs a preconfigured client ID.
• Default OAuth apps cover Claude, Claude Code, ChatGPT, Codex and Cursor, so no admin has to create one. An account admin can turn them on or off.
• Users need the Consumer access entitlement and permission to call the specific MCP.

What can a Databricks MCP server read and change?

It depends on the server. Genie One asks questions and returns answers, and Genie Agent doesn't write to your tables. Databricks SQL permits reads and writes by default, within your Unity Catalog permissions. Set disallow_writes to true in system.ai.dbsql_policy to make it read-only.

What it can read

• Genie One: genie_ask, genie_poll_response, genie_get_query_result, genie_cancel_response and view_ask. Unity Catalog permissions are always enforced.
• Genie Agent: plain-language questions against one Genie Agent.
• Databricks SQL: execute_sql on a SQL warehouse, with the caller's permissions.

What it can change

• Databricks SQL: execute_sql writes as well as reads until you set disallow_writes.
• system.ai.sandbox: runs Python, SQL or shell code in an isolated environment with no network egress.
• system.ai.github: read-only on public repositories by default. Private repositories and writes need the Databricks GitHub connector app.

How do you set up a Databricks MCP server in Claude, ChatGPT and Cursor?

Databricks documents all three. Genie plugins connect Claude and ChatGPT to Genie One with default OAuth apps, so users sign in with a Databricks account. Cursor uses the Unity Gateway CLI or ~/.cursor/mcp.json. For other MCPs in Claude or ChatGPT, add a custom server with your workspace URL.

{
  "mcpServers": {
    "databricks-genie": {
      "url": "https://<workspace-hostname>/ai-gateway/mcp-services/system.ai.genie_one_mcp?o=<workspace-id>",
      "headers": {
        "Authorization": "Bearer <databricks-pat>"
      }
    }
  }
}

What can you ask a Databricks MCP server?

Genie One suits business questions, because it resolves terms through Genie Ontology before it writes SQL. These examples use its documented tools and Databricks SQL's execute_sql. Keep Databricks SQL read-only until you've set up approvals, because it permits writes by default. The last example writes to your workspace.

• Reads

  What was paid media spend by channel last quarter, and which channel's cost per lead rose the most? Link the sources you used.

  Copy

• Reads

  Give me the full result behind that answer, with every campaign and its spend.

  Copy

• Reads

  Using Databricks SQL, list the tables in marketing.paid_media with their row counts. Run SELECT statements only.

  Copy

• Writes

  Create a view in marketing.reporting that joins campaign spend to CRM opportunities. Show me the SQL first and wait for my approval.

  Copy

What are the limits of a Databricks MCP server?

Genie One truncates results in genie_ask and genie_poll_response, so agents call genie_get_query_result for full rows. Databricks SQL truncates large results too. A Genie Agent covers up to 50 Unity Catalog objects. Databricks cites a 90-second SQL execution timeout and a workspace Genie rate limit.

• Calls to workspace endpoints skip Unity Gateway, so they miss its tracing, MCP grants, policies and guardrails.
• The Genie Agent server passes no conversation history to the Genie API.
• system.ai.web_search and system.ai.sandbox aren't available to users with only consumer access.
• If your workspace uses IP access lists, allow ChatGPT's outbound IPs.

How much does a Databricks MCP server cost?

Databricks lists no separate MCP fee. Genie One and Genie Agents usage by users is free until 31 January 2027, and service principal usage is billed. Genie compute such as SQL Serverless is billed separately. Genie Agent MCP calls use Databricks SQL pricing, and other servers follow their own pricing.

• Each user also gets 150 DBUs a month free, which Databricks values at $10.50 in US East.
• Unity Catalog functions use serverless general compute pricing, and AI Search indexes use AI Search pricing.

What are the alternatives to a Databricks MCP server?

The main alternatives are a rival warehouse's server, dbt for modeled metrics, or a BI server on top of Databricks. Pick the layer that holds your trusted numbers. Marketing analytics MCP servers compares warehouse servers with reporting connectors and direct platform servers for marketing questions.

• Snowflake MCP: The warehouse route, if your data sits in Snowflake.
• BigQuery MCP: The warehouse route, if your data sits in BigQuery.
• dbt MCP: The route if your metrics are defined in dbt models.
• Power BI MCP: Reads reports and models, if Power BI sits on top of Databricks.

Which skill goes with a Databricks MCP server?

Use the free Agent write safety skill before you connect Databricks SQL, because it permits writes by default and an agent could change live tables. Use the Marketing context kit to write down your table names and metric definitions once, so every answer starts from the same rules and the assistant has less to guess.

• Agent write safety: Sets approvals before an agent can run SQL that changes data.
• Marketing context kit: Holds your table names and metric definitions in one folder.