Trodo

generateText, streamText, generateObject, and tool calls capture automatically on v7, or with experimental_telemetry on v5/v6.

Every ai call inside wrap your agent becomes a span, model calls as llm, tool executions as tool.

How the AI SDK reports telemetry changed in v7, so which setup you need depends on your version:

AI SDK	What you do
v7+	Nothing, on `trodo-node` 2.23.4 or later. `registerOTel()` registers Trodo's telemetry integration for you, including under ESM and Next.js.
v5 / v6	Pass `experimental_telemetry: { isEnabled: true }` per call.

Note

How v7 differs. Through v6 the AI SDK emitted OpenTelemetry spans itself, so any OTel pipeline picked them up. v7 replaced that with a telemetry registry: registerTelemetry(...) takes objects of callbacks, and a call with experimental_telemetry: { isEnabled: true } but no registered integration emits nothing at all.

registerOTel() registers Trodo's integration when it finds ai installed, so v7 works with no extra package and no per-call flag. To wire it yourself, say you configure the registry in one place, the integration is exported:

import trodo from 'trodo-node';
import { registerTelemetry } from 'ai';

registerTelemetry(trodo.aiSdkTelemetry());

Do that instead of letting registerOTel() do it (pass disableInstrumentations: ['vercel-ai'] to trodo.init()), not as well, the AI SDK appends every integration it is given, so registering twice duplicates every span.

On trodo-node 2.23.3 and earlier, v7 tool spans carry no output, and under ESM or Next.js the integration is not registered at all. Upgrade to 2.23.4.

Registering @ai-sdk/otel alongside this also double-counts: it routes the same events into OpenTelemetry, which Trodo already consumes. Pick one.

What's captured

Call	Span kind	Auto-extracted
`generateText`	`llm`	model, provider, tokens, prompt, completion
`streamText`	`llm`	same, accumulated
`generateObject` / `streamObject`	`llm`	model, tokens, schema, parsed object
`tool()` and the tool-call loop	`tool`	tool name, input, output
`embed` / `embedMany`	`llm`	model, tokens

The provider field is whichever adapter you use (openai, anthropic, google).

Install

npm install ai @ai-sdk/openai

No OpenTelemetry package is needed on any version.

Minimal example

import trodo from 'trodo-node';
import { generateText, tool } from 'ai';
import { openai } from '@ai-sdk/openai';
import { z } from 'zod';

trodo.init({ siteId: process.env.TRODO_SITE_ID });

await trodo.wrapAgent('ai-sdk-bot', async (run) => {
  const { text } = await generateText({
    model: openai('gpt-4o-mini'),
    prompt: 'What is the weather in SF?',
    tools: {
      get_weather: tool({
        description: 'Current weather in a city',
        inputSchema: z.object({ city: z.string() }), // AI SDK v5+: inputSchema (was `parameters` in v4)
        execute: async ({ city }) => `Sunny in ${city}, 72F`,
      }),
    },
    experimental_telemetry: { isEnabled: true }, // required
  });
  run.setOutput(text);
});

Warning

Streaming reports no tokens unless the provider is asked for them. With streamText, usage arrives in the final chunk only if the request opted in. @ai-sdk/openai-compatible does not by default, so the AI SDK itself receives none, await result.usage resolves empty, and no telemetry integration can report what the SDK never got. The span is still recorded; its token counts and cost are simply absent.

const provider = createOpenAICompatible({
  name: 'my-gateway',
  baseURL: '…',
  includeUsage: true,   // ← without this, streamed calls report 0 tokens
});

Measured against ai 7.0.79: 0 tokens by default, correct counts with includeUsage: true. Provider configuration, not tracing, non-streaming calls are unaffected, and other providers have their own equivalent option.

On v5 / v6 the experimental_telemetry: { isEnabled: true } flag is required, without it the AI SDK emits no spans. On v7 it is harmless to leave in, and harmless to remove.

The provider on each span is the vendor alone (openai, anthropic, google), so cost and provider filters line up with every other span. The AI SDK's fuller id (openai.responses, anthropic.messages) is kept on the span as ai.provider_id, alongside ai.finish_reason.

Vercel AI SDK

What's captured

Install

Minimal example

On this page