Databricks MCP server: ask questions, get governed answers

Databricks AI analytics: what you can ask

Typical questions from data and analytics teams. Each answer shows the definition, source and query behind it.

• What was net revenue by region last quarter, using our approved revenue definition?
• Which customers had the largest drop in usage versus the prior month?
• Show daily order counts for the last 60 days.
• Which product lines had the largest margin change month over month?
• What are our top 10 products by units sold this year?
• How many active customers do we have by segment compared to last year?

Worked example: a governed Databricks answer

One question, and everything the answer shows so you can check it. The figures are invented sample data.

SELECT cur.account_name,
       prev.active_users AS prior_month,
       cur.active_users  AS last_month,
       cur.active_users - prev.active_users AS change
FROM main.gold.account_usage_monthly cur
JOIN main.gold.account_usage_monthly prev
  ON prev.account_id = cur.account_id
 AND prev.month = add_months(cur.month, -1)
WHERE cur.month = add_months(date_trunc('MM', current_date()), -1)
ORDER BY change
LIMIT 3;

Common pitfalls when asking Databricks questions

Each one is settled once in an approved definition, so every person and every assistant gets the same answer.

• Bronze, silver and gold copies

  The same events often exist raw, cleaned and aggregated. The definition names the gold table, so answers do not drift between layers.

• Same table name in different catalogs

  Unity Catalog names have three parts: catalog, schema, table. A dev and a prod catalog can both hold usage_monthly; the definition uses the full name.

• Late-arriving data

  Events that land late change last month's totals after the fact. The definition states when a month is treated as final, so a re-asked question is explainable.

How Nexatron connects to Databricks

Whose permissions

Questions run through the connection's Databricks account, so its grants are the outer limit. Inside that, Nexatron roles and the row policies your administrator defines decide what each person can see.

Read and write

Read-only. Nexatron reads Databricks data to answer questions and never writes to Databricks.

Setup

An administrator connects a service credential (API token or service-account key) issued by the vendor. Steps: choose the source → check prerequisites → enter connection details → verify → ask. No data migration and no pipeline to build: data is queried where it lives.

Databricks MCP connector for Claude, ChatGPT and Copilot

Nexatron's hosted MCP server is a governed Databricks MCP connector: Claude, ChatGPT and Microsoft Copilot ask questions of your Databricks data through it. Each person signs in as themselves, so the assistant sees only what they may see.

Connect Databricks to Claude, ChatGPT or Copilot

1. Connect Databricks to your Nexatron workspace.
2. In Claude, add Nexatron from Claude's connector directory. In ChatGPT or Copilot, add a custom MCP connector with the server URL https://api.nexatron.io/api/v1/mcp.
3. Sign in with your own Nexatron identity (OAuth 2.1), then ask your Databricks questions.

Databricks connector FAQ

Can I ask Databricks questions without writing queries?

Yes. The connector is available now. Ask in Teams, Slack, the Nexatron web app, or an AI assistant connected through Nexatron's MCP server.

Does Nexatron respect Databricks permissions?

Yes. The Databricks account's grants are the outer limit, and your row policies narrow it for each person.

Can Nexatron change data in Databricks?

No. Nexatron only reads Databricks data.

What do I need to connect Databricks?

A vendor-issued service credential and an administrator to connect it.

Is there a Databricks MCP server for Claude, ChatGPT or Copilot?

Yes: Nexatron's hosted MCP server, once Databricks is connected to your workspace. It is Nexatron's server, not one published by Databricks, and the assistant gets the same governed answer as the Nexatron app.

How is Nexatron different from Databricks' own AI tools?

Databricks' AI tools work inside the Databricks workspace. Nexatron brings governed answers to Teams, Slack, the web app and connected assistants, uses one set of approved definitions across all your systems, and shows the SQL behind each answer.

Can business users ask Databricks questions without SQL or a notebook?

Yes. They ask and get the number with its definition and SQL. Analysts keep their notebooks and dashboards; Nexatron gives everyone else a governed way to ask.

Which Databricks identity runs the queries?

An administrator connects one service credential for the workspace, and its Unity Catalog grants are the outer limit. Inside that, Nexatron roles and row policies decide what each person can see.