Connect Databricks to Claude, ChatGPT and Copilot
Databricks is a lakehouse platform where data lands in Delta tables governed by Unity Catalog. Nexatron runs each question as SQL on a Databricks SQL warehouse your administrator chooses, under the asking person's own Databricks identity.
Nexatron's hosted MCP server is listed in Claude's connector directory. ChatGPT and Microsoft 365 Copilot add the same server as a custom MCP connector. Connect Databricks to your Nexatron workspace once, then ask from whichever assistant you use.
What you can ask
- Which accounts had the largest drop in active users versus the prior month?
- What was contribution margin by product line last month, from the gold finance tables?
- Which plants had the most unplanned downtime hours this quarter?
- How many active devices reported telemetry each day last week?
- Which SQL warehouses used the most DBUs last month, by team?
- What is the churn rate of customers who opened more than three support tickets?
Questions read Unity Catalog's catalogs, schemas, tables, views and materialized views, and system tables such
as system.billing.usage where the person may read them. Queries are read-only SELECTs, visible in the
warehouse's query history.
Prerequisites
- A Nexatron workspace admin to add the connection.
- A Databricks SQL warehouse, and a setup token or service principal that can read the catalog in scope.
- Access to the Databricks account console to add an app connection, for people's OAuth sign-in.
- A Databricks identity for each person who asks questions, with Unity Catalog grants on what they should query.
Admin setup
- In Nexatron, go to Knowledge ▸ Sources ▸ New connection and choose Databricks. Enter the server hostname and the SQL warehouse's HTTP path (SQL warehouse ▸ Connection details ▸ HTTP path), the setup access token, and optionally a catalog and schema. The token is the connection's setup role: it reads metadata only, for setup, and never answers anyone's question.
- Open the connection's page. Under Sign-in for each person, copy the redirect URI.
- In the Databricks account console ▸ Settings ▸ App connections, add a connection with that redirect URI,
the scopes
sql,offline_access,openidandemail, and a client secret. - Paste the client ID and client secret into Nexatron and select Save client. Replacing the client later ends everyone's sign-ins.
- Choose who can use the connection.
How each person signs in
Each person opens Knowledge ▸ Sources ▸ Your sign-ins and signs in with Databricks, on the Databricks workspace's own page. Nexatron never sees their password. Instead, a person can use their own personal access token: enter their user name and token under Or use a personal access token, and select Link my login.
First question to try
After you add Nexatron in Claude, ChatGPT or Copilot, ask:
Which accounts had the largest drop in active users versus the prior month?
The answer shows the tables and SQL, so a data engineer can open the same statement in the SQL editor.
Permissions and masking
Questions run as the person's own Databricks identity, so Unity Catalog grants, row filters and column masks apply before results return.
Nexatron's own rules under Settings ▸ Workspace ▸ Access (Access) apply on top: row rules limit rows and column masking hides a column's values. They narrow what Databricks allows and never widen it. A question that would total, average or count a masked column is refused in words, and an answer that ran under rules says so in its evidence.
Troubleshooting
| Symptom | Cause and fix |
|---|---|
| Signing in with Databricks isn't set up for the connection yet | The admin has not saved the app connection's client under Sign-in for each person on the connection's page. |
| Databricks rejects the redirect URI | The app connection's redirect URI must be the one shown on the connection's page, exactly. |
| Your login for the connection stopped working | Your personal access token expired or was revoked. Select Replace login, or sign in with Databricks. |
| An answer is refused for access | Your Unity Catalog grants do not cover a table the question needs. |
| A question about a masked column's total is refused | A Nexatron masking rule covers that column, and aggregates over masked columns are always refused. |
Related
- Databricks on nexatron.io: what the connector covers
- Connect Claude: add Nexatron from Claude's connector directory
- Connect ChatGPT: add Nexatron as a custom connector
- Connect Microsoft 365 Copilot: a declarative agent or a Copilot Studio tool
- MCP reference: tools, OAuth scopes and errors