Vercel AI SDK

The Vercel AI SDK (the ai npm package) supports OpenTelemetry through its official @ai-sdk/otel integration. Attach that integration to a native ToolLoopAgent and configure the Logfire TypeScript SDK as the global tracer provider to send the resulting agent, model, and tool spans to Logfire.

Installation

Terminal

npm install 'ai@^7' '@ai-sdk/openai@^4' '@ai-sdk/otel@^1' @pydantic/logfire-node zod

Usage

Configure Logfire first so the global tracer exists before any ai call:

agent.ts

import * as logfire from '@pydantic/logfire-node';
import { ToolLoopAgent, stepCountIs, tool } from 'ai';
import { OpenTelemetry } from '@ai-sdk/otel';
import { openai } from '@ai-sdk/openai';
import { z } from 'zod';

logfire.configure({ serviceName: 'vercel-ai-agent' }); // sets the global OTel tracer provider

let toolCalls = 0;
const weather = tool({
  description: 'Get the weather for a city',
  inputSchema: z.object({ city: z.string() }),
  execute: async ({ city }) => {
    toolCalls += 1;
    return { city, tempC: 21 };
  },
});

const agent = new ToolLoopAgent({
  model: openai('gpt-4o-mini'),
  instructions: 'Use the weather tool before answering.',
  tools: { weather },
  stopWhen: stepCountIs(3),
  telemetry: {
    functionId: 'weather-agent',
    integrations: [new OpenTelemetry()],
    isEnabled: true,
  },
});

async function main() {
  try {
    const { text } = await agent.generate({
      prompt: 'What is the weather in Paris? Use the tool.',
    });
    if (toolCalls !== 1) throw new Error(`Expected one tool call, received ${toolCalls}`);
    console.log(text);
  } finally {
    await logfire.shutdown(); // flush success and error spans before exit
  }
}

main();

Set your OPENAI_API_KEY and LOGFIRE_TOKEN, then run with npx tsx agent.ts. The example fails unless the native ToolLoopAgent executes weather. You’ll see spans for the agent, prompt, response, token counts, and tool call in Logfire. Vercel AI SDK runs also appear in the specialized Agents view; the support matrix shows which columns each view populates.

Common pitfalls

• Register both pieces before running the agent. Configure Logfire’s global tracer provider and attach new OpenTelemetry() in the agent’s telemetry.integrations.
• Flush on exit. Short-lived scripts must await logfire.shutdown() or the batch processor drops unflushed spans.
• Content capture. recordInputs / recordOutputs default to true, so prompts and completions are captured. Set them to false in telemetry to redact sensitive content.

Next.js

Install the Next.js OpenTelemetry helper with npm install @vercel/otel. In Next.js, call registerOTel(...) from @vercel/otel and registerTelemetry(new OpenTelemetry()) from ai and @ai-sdk/otel inside the exported register() function in instrumentation.ts. Point the OpenTelemetry Protocol (OTLP) exporter at Logfire with OTEL_EXPORTER_OTLP_ENDPOINT=https://logfire-us.pydantic.dev and OTEL_EXPORTER_OTLP_HEADERS='Authorization=<write-token>'.

Managed prompts

Author and version prompts in Prompt Management and fetch them with the Logfire TypeScript SDK:

import { defineTemplateVar } from '@pydantic/logfire-node/vars';

const promptVar = defineTemplateVar<string, { city: string }>('prompt__weather', {
  default: 'What is the weather in {{city}}? Use the tool.',
  templateInputsSchema: {
    type: 'object',
    properties: { city: { type: 'string' } },
    required: ['city'],
  },
});

const resolved = await promptVar.get({ city: 'Paris' });
// Pass resolved.value as the `prompt` to agent.generate().

See Use Prompts in Your Application for the full workflow.