Vercel AI SDK

The Vercel AI SDK is the one framework here that is not OpenTelemetry-native, so it does not use the exporter-only recipe the others do. It has its own telemetry API, and Tracelane ships an integration for it.

Why this page is different. As of ai@7, generateText emits no OpenTelemetry spans by itself — the package has no @opentelemetry/api dependency. Passing experimental_telemetry: { isEnabled: true } alone gets you nothing. The AI SDK instead exposes registerTelemetry(), and tracelaneTelemetry() maps it onto OpenTelemetry.

What you get


Agent structure	Yes — one span per operation, per step, per model call and per tool call, correctly nested.
The LLM leg	Yes, with token counts and finish reason, inside the tree.
Multi-step tool loops	Yes — each step is its own span, so a tool-calling loop is visible as a tree rather than one flat call.

Setup

npm install @tracelanedev/sdk \
            @opentelemetry/sdk-trace-node \
            @opentelemetry/exporter-trace-otlp-proto

import { NodeTracerProvider, BatchSpanProcessor } from "@opentelemetry/sdk-trace-node";
import { OTLPTraceExporter } from "@opentelemetry/exporter-trace-otlp-proto";
import { resourceFromAttributes } from "@opentelemetry/resources";
import { registerTelemetry } from "ai";
import { tracelaneTelemetry } from "@tracelanedev/sdk/vercel_ai";

const provider = new NodeTracerProvider({
  resource: resourceFromAttributes({ "service.name": "my-agent" }),
  spanProcessors: [
    new BatchSpanProcessor(
      new OTLPTraceExporter({
        url: "https://gateway.tracelane.dev/v1/traces",
        headers: { "x-tracelane-api-key": process.env.TRACELANE_API_KEY! },
      }),
    ),
  ],
});
provider.register();

registerTelemetry(tracelaneTelemetry());

Then enable telemetry per call — the AI SDK requires this, and Tracelane cannot turn it on for you:

await generateText({
  model,
  prompt: "…",
  experimental_telemetry: { isEnabled: true, functionId: "my-agent" },
});

Your API key needs the ingest scope. Mint one in Settings → API keys.

Flush before your process exits. BatchSpanProcessor buffers, so a short-lived script or a serverless invocation can exit with spans still in memory and nothing will arrive. Call await provider.shutdown() before exit. This is the single most common reason a correct setup shows no traces.

What a trace looks like

A two-step tool-calling run, verified end to end against Tracelane production on 2026-08-31 — 6 spans, 1 root, 5 parented, 0 orphans, read back from storage:

ai.generateText                    (root)
├── ai.step 0
│   ├── ai.languageModelCall
│   └── ai.toolCall lookup
└── ai.step 1
    └── ai.languageModelCall

Span volume

An OTel span and a gateway call are not the same unit, but they count against the same meter. A gateway request and an OTel span both count as ingest bytes — there is no separate quota for one path or the other. On this path your app sends spans, and one agent run is many — the trace above is 6 for a single generateText, and each extra step and tool call adds more. Size your ingest estimate against spans per run, not runs, and see pricing.

The alternative: the gateway base URL

If you only need the model call and want routing, BYOK, guardrails, caching and budgets, point the SDK’s provider at the Tracelane gateway instead:

import { createOpenAI } from "@ai-sdk/openai";

const openai = createOpenAI({
  baseURL: "https://gateway.tracelane.dev/v1",
  apiKey: process.env.TRACELANE_API_KEY,
});

	Gateway base URL	`registerTelemetry` (above)
Captures the model call	yes	yes
Captures steps and tool calls	no	yes
Trace shape	one span per request	a nested tree
Routing, BYOK, guardrails, caching	yes	no — calls go direct to your provider

They compose: run both and you get gateway features plus agent structure. Note the two do not currently join into a single trace — a gateway-proxied call does not adopt your agent’s trace context, so it appears as its own single-span trace.

Earlier releases exported instrumentVercelAI(ai). It never worked — it mutated the imported ai module object, and ES module namespaces are read-only. It now throws an error naming this replacement.

Get started

SDKs

Integrations

Concepts

Features

Tools

Audit & compliance

Operations

Migrations

Comparisons

What you get

Setup

What a trace looks like

Span volume

The alternative: the gateway base URL

​What you get

​Setup

​What a trace looks like

​Span volume

​The alternative: the gateway base URL

What you get

Setup

What a trace looks like

Span volume

The alternative: the gateway base URL