Observability for your Vercel AI SDK app.
The Vercel AI SDK powers streaming AI in your TypeScript app; Obsivara tells you what each call costs and how it performs. Every generateText, streamText, and tool call traced with tokens, latency, and dollars — across providers, so switching models is a measured decision. The SDK has first-class OpenTelemetry: flip on experimental_telemetry and point your exporter at Obsivara. Your app calls providers directly, out of the request path.
Vercel AI SDK, fully observable.
Span-level call tracing
Every generateText, streamText, generateObject, and tool call captured as a span with inputs, outputs, tokens, latency, and the provider and model used.
Cross-provider cost attribution
The SDK's unified provider layer makes model-switching easy; Obsivara prices each call so you see cost per route, model, and feature across every provider you use.
Streaming and tool visibility
Time-to-first-token and full-stream latency for streamed responses, plus tool-call results and durations — so slow or failing UX points to the exact step.
Failure attribution and health
Errors, timeouts, and failed tool calls classified by provider and route, with health scoring and predictive alerts.
Three steps. No code changes.
Turn on experimental_telemetry
The Vercel AI SDK emits OpenTelemetry via its experimental_telemetry option — enable it and point your OTLP exporter at Obsivara (or use the Obsivara SDK). Your app keeps calling providers directly; Obsivara records each call out-of-band, with no proxy in the path and no added latency.
Run your app
Traces flow as your routes and server actions execute — no changes to your generateText or streamText logic.
See it in Obsivara
Routes and models appear in the inventory and knowledge map with per-call cost, latency, and health within minutes.
Signals tracked out of the box.
- →Call spans per route and model
- →Tokens and cost per provider and feature
- →Time-to-first-token and stream latency
- →Tool calls and error rates
- →Model and provider switches over time
- →Health score and predictive alerts
Common questions.
Does Obsivara use the AI SDK's built-in telemetry?
Yes. The SDK emits OpenTelemetry through its experimental_telemetry option; Obsivara ingests those spans over OTLP, so you don't instrument your calls by hand.
Does it work across the providers the AI SDK supports?
Yes. Because the SDK normalizes OpenAI, Anthropic, Google, and others behind one interface, Obsivara prices and compares calls across every provider you route to.
Does it capture streaming responses?
Yes. Streamed calls are traced with time-to-first-token and full-stream latency, so streaming UX problems are visible per route.