Databricks adds native Claude OpenTelemetry export to MLflow

Databricks made native OpenTelemetry export from Claude Platform to MLflow available in public beta on Nov. 3. The integration gives developers a direct route to send execution traces from Anthropic SDKs, Messages API requests and Claude Managed Agents sessions into an MLflow experiment.

Claude Platform sends OpenTelemetry spans—individual steps such as model requests and tool uses—over the OpenTelemetry Protocol. Databricks says Zerobus Ingest receives those spans and stores them in Unity Catalog tables linked to an MLflow experiment. The company presents the setup as a way to apply Unity Catalog access controls, query traces with SQL and use the records for debugging and evaluation.

The workflow has three steps, according to Databricks: create an MLflow experiment with Unity Catalog selected as the trace location, configure the Databricks destination or exporter, then run a Claude task and inspect its trace. Its Unity Catalog tracing documentation lists prerequisites including a Unity Catalog-enabled workspace, a supported region, a SQL warehouse with the required permission, MLflow 3.14 or later, and explicit catalog and table privileges. The Zerobus documentation says the OTLP endpoint writes telemetry to pre-created Unity Catalog Delta tables and requires table-scoped authorization.

Anthropic’s OpenTelemetry documentation also labels the tracing capability as beta and lists Databricks as a supported partner destination. Metadata-only collection is the default, Anthropic says. Capturing prompts and responses requires opting into Full content, after which the destination stores that material under its own access and retention settings.

Once traces reach MLflow, Databricks says users can inspect model calls, tool activity and timing, send cases for expert review, and compare evaluations after changing an agent. The company also says MLflow Issues can identify and group recurring quality or operational problems across traces. Databricks labels MLflow Issues as beta; its announcement provides no independent performance measurements for either the feature or the integration.