Skip to main content
Serve agent inference from your Databricks workspace’s model serving endpoints — foundation models (pay-per-token), provisioned throughput, or external-model gateways. Credentials come from an existing Databricks data connection; no separate key is stored.

​
Prerequisites

A Databricks data connection authenticated with a personal access token, M2M OAuth, or Azure Entra ID service principal. (Per-user OAuth connections can’t be used for inference — the agent runs without a user context.) That connection’s service principal also needs Can Query on the endpoints you expose — a separate ACL from its warehouse and catalog grants — and the workspace’s AI Gateway throughput limit raised above its low default. See Databricks AI for migrations for both steps.

​
Configure in Datafold

  1. Navigate to Settings → Integrations, click Add New Integration, and choose Databricks.
  2. Select the Data connection whose workspace hosts your serving endpoints.
  3. Models — the workspace’s ready LLM serving endpoints are listed automatically. Select the ones to expose (e.g. databricks-claude-sonnet-4-5).
  4. Model API surfaces — Datafold infers each endpoint’s wire protocol (Anthropic messages for Claude endpoints, chat completions or Responses for others) from its name. Override per model for custom-named endpoints.
  5. Click Save.