Developer questions, answered.

No. A base URL, an OAuth token and headers are enough. The optional intsen package only helps with golden tests in CI.

No. It stays in your environment. IntSen sees the model calls it routes through the gateway.

Calls still work and are recorded, but each call becomes its own turn. The Console warns that the task id is missing.

The model answers without asking for a tool and no new call with that task id follows for 20 seconds. Turns with no activity are also closed after an idle timeout.

IntSen detects a new user message after a final answer and starts a new turn segment, but a new task id per user message gives the clearest results.

Yes. Checks on streamed answers run after completion, or during the stream if you switch that on. Tool calls are assembled from streamed deltas.

Request records with cost, timings and verdicts. Request and response bodies only if capture is on, for your retention period, with personal data redacted if you choose. Guardrail results store categories, never content.

You still get tracing with tool names, call ids and timings, plus cost and verdicts—but no arguments, results or content, and evaluations that need content cannot run.

Model-based checks and judges run on your registration with the model you choose. Each call is tagged, so the Console shows them separately from your agent's own cost.

OpenAI in the first release: chat completions, Responses and embeddings, with streaming. Azure OpenAI and Anthropic are planned.

Yes. intsenctl apply --dry-run validates and shows the diff before anything changes, and golden tests in CI catch quality regressions.

You get an OpenAI-compatible error body with a stable code (for example model_not_allowed or task_budget_exceeded) and a request id to look up in the Console.

Technically, any code holding a provider key can call the provider directly, and those calls are not governed. That's why the gateway holds the provider keys: removing direct keys from teams is part of rollout.