@mlflow/typesafe

Yes. The documented quickstart begins by starting an MLflow Tracking Server, then initializes MLflow once when the application starts using mlflow.init with a trackingUri and experimentId. Without that server and initialization there is nothing for the traced calls to report to.

The integration traces systemOne() calls and captures their state, questions, structured answers, model, token usage, latency, request ID, and errors. Transport options such as API keys, headers, timeouts, retries, and abort signals are explicitly not recorded.

No. Only systemOne() calls are traced. Other SDK operations, including models.list(), are not traced, so anything you need observed must run through the wrapped client's systemOne path.

Import tracedTypeSafe from @mlflow/typesafe and the TypeSafeClient from @typesafe-ai/sdk, then pass a new client into tracedTypeSafe. Call systemOne() on the returned client as usual; state, questions, and structured answers are sent to MLflow without changing the call sites.

The project is licensed under the Apache License 2.0, as stated in the package documentation for the MLflow TypeScript SDK TypeSafe AI integration.

Skip it if you do not run an MLflow Tracking Server, if you need tracing for SDK operations other than systemOne() such as models.list(), or if you must capture transport-level details like API keys, headers, timeouts, retries, or abort signals, which the integration does not record.