Respan

Respan is an AI gateway with observability, evaluations, and security testing. Its tracing tools give developers a way to inspect the steps inside an AI request, including model calls, tool execution, and agent activity, instead of seeing only the final answer or error.

Follow a request through its spans

The tracing quickstart describes each recorded operation as a span. A trace groups the spans from one request or workflow into a tree; a thread connects spans across related turns. This structure helps you locate which operation consumed time or tokens and inspect its inputs and outputs.

Respan's SDK traces supported model-provider SDKs after initialization. The documented integrations differ between Python and TypeScript, so confirm support for the library your application uses rather than assuming all providers behave alike. A workflow wrapper groups related calls into one trace.

That trace-based investigation has a narrower AI focus than BitDive's Java runtime capture, which records method calls, database activity, and other application behavior. The distinction matters if the failure occurs outside the model request: an LLM span may show a slow tool call without explaining every internal operation of the service behind it.

Initialize once and check delivery

The quickstart instructs developers to initialize the SDK before the first model call. It distinguishes Production and Test API keys, and the Logs interface can filter those environments. An apparently missing trace may belong to a different environment or time range rather than a failed request.

The SDK sends spans in the background and flushes on process exit. For serverless functions, the documentation requires an explicit flush before returning. Review that lifecycle requirement when integrating a short-lived handler, where finishing the business operation does not necessarily mean the trace has reached the dashboard.

Gateway and security features need their own evaluation

Respan advertises access to more than 1,000 models through one API, PII masking, and agent red-teaming. It also names span-01 for scoring agent behavior and span-router for selecting a model per request. Those are product capabilities to evaluate against your workload, not guarantees of safe output or interchangeable model behavior.

Start with a free account as directed by the documentation, then check current pricing and data-handling terms. Traces can contain prompts, outputs, and tool inputs. Decide what must be excluded or masked before sending customer material, and verify that the collected information is sufficient to explain a failure without exposing unnecessary sensitive data.