Vercel AI SDK

About the Vercel AI SDK integration

@telemetry-dev/ai-sdk plugs into the AI SDK's telemetry surface. On ai@7 it covers generateText, streamText, Agent, generateObject, streamObject, embed, embedMany, and rerank; a dedicated @telemetry-dev/ai-sdk/v6 entry covers ai@6. Each call becomes its own OpenTelemetry trace: a root span, a chat span per model step, and an execute_tool span per tool call — with provider requests and tool execute runs parented into the same trace.

Key features

• One integration entry: Add telemetryDev() to telemetry.integrations (or experimental_telemetry on v6) — no manual spans, no exporter config.
• Agent loops and tools: Multi-step runs and nested AI SDK calls made from inside a tool stay in the outer call's trace.
• Tokens and cost: Cache-read, cache-write, and reasoning token breakdowns when the provider reports them; cost computed server-side from real usage and current pricing.
• Explicit user context: runtimeContext keys reach telemetry only when opted in via includeRuntimeContext — userId maps to user.id, sessionId to the conversation id.
• Standard OTLP, no lock-in: Built on OpenTelemetry, so your data stays portable.

Get started

import { generateText } from "ai";
import { telemetryDev } from "@telemetry-dev/ai-sdk";

const { text } = await generateText({
  model,
  prompt: "Summarize the incident report.",
  runtimeContext: { userId: "u_123", sessionId: "s_456" },
  telemetry: {
    functionId: "summarize-incident",
    includeRuntimeContext: { userId: true, sessionId: true },
    integrations: [telemetryDev()],
  },
});

Set TELEMETRY_DEV_API_KEY and your first traces land in about five minutes — no key means the integration is a complete no-op. Requires ai >= 6.0.111; import from the entry matching your major version.