How to Connect Claude to Databricks

Databricks is increasingly where companies land when they need one platform for data engineering, ML, and analytics. Connecting it to Claude is now straightforward: Databricks ships its own MCP server for Claude. As with every warehouse, the connection is the easy part. Making the answers right for a whole team is where the three options split.

Your three options

Option 1: Genie One MCP (native)

Databricks' business-user product, Genie One, is now reachable from Claude through a managed MCP server. It is generally available and documented for both Claude and ChatGPT.^[1] You add https://<workspace>/ai-gateway/mcp-services/system.ai.genie_one_mcp as a connector and sign in with OAuth; Databricks recommends on-behalf-of-user auth, so Unity Catalog permissions apply per person. If you set up the earlier beta endpoint, move off it: it is deprecated on October 31, 2026.^[1]

Genie One answers through Genie Ontology, Databricks' governed context layer, and the curated scopes underneath it are Genie agents (renamed from Genie spaces in July 2026). Each agent holds up to 50 tables, though Databricks' own guidance is five or fewer for accuracy, and the docs expect an analyst who knows SQL to keep adding context as people use it.^[2][3] Data outside Databricks is reachable through Lakehouse Federation, once someone sets up the connections in Unity Catalog.^[4]

User usage is free through January 31, 2027. The SQL warehouse compute that runs the queries is billed the whole time, and so is service principal usage, which is how many API and MCP agents authenticate.^[5]

It is the right pick if your data already lives in Unity Catalog and you have people whose job includes curating agents.

Option 2: A DIY MCP server

The other route is pointing a general-purpose MCP server, or Databricks' lower-level managed MCP endpoints,^[6] at your workspace with a token. Claude can then list tables and run SQL.

• No business context: Claude sees column names, not what they mean. "Revenue" still has to be defined somewhere.
• Per-person setup: each team member configures their own connection, and nothing is managed centrally.
• Access is the credential's: whoever's token it is, Claude sees what they see.

It works for a data engineer exploring a schema they already know.

Option 3: Databricks plus Claude through Contextflo

Contextflo connects to your Databricks workspace once, read-only, generates context for your tables, and serves it to Claude and ChatGPT over MCP. Your team asks in the AI they already use, with shared definitions and per-person table access, and Databricks can sit next to your other sources.

1. Connect Databricks. In Contextflo, add a Databricks data source with your workspace URL and a service principal or personal access token that has read access.
2. Sync your tables. Pick the catalogs, schemas, and tables to include. Contextflo generates descriptions for each table and column from Unity Catalog metadata, plus any code and docs you connect.
3. Connect Claude. In Claude, add Contextflo as a connector with https://mcp.contextflo.com/mcp and sign in. Each teammate connects once and gets the same definitions.
4. Ask your first question. For example: "What was net revenue last month, by region?" Claude writes the SQL, runs it on your warehouse, and shows the query and the definition it used.

Which to choose

Genie One MCP if you are all-in on Databricks and have a data team to curate agents. You get row-level governance from Unity Catalog and near-zero license cost until 2027.

A DIY MCP server for one technical person who knows the schema and wants raw access.

Contextflo if non-technical people need answers in Claude or ChatGPT, your data lives in more than one place, and nobody's job is curating a semantic layer.

Here is a more in-depth look at Contextflo and how it works.