Vercel AI SDK

axonPushMiddleware returns a LanguageModelMiddleware for the Vercel AI SDK. Wrap a model with it once and every call through that model is traced, the call sites themselves do not change.

Tested against ai@^4.

Install

npm install @axonpush/sdk ai

Wrap the model

import { wrapLanguageModel } from "ai";
import { openai } from "@ai-sdk/openai";
import { AxonPush, axonPushMiddleware } from "@axonpush/sdk";

const client = new AxonPush();

const model = wrapLanguageModel({
  model: openai("gpt-4o"),
  middleware: axonPushMiddleware({
    client,
    channelId: process.env.AXONPUSH_CHANNEL_ID!,
    agentId: "support-bot",
  }),
});

Then use the wrapped model everywhere:

import { generateText, streamText } from "ai";

const { text } = await generateText({ model, prompt: "Explain quantum computing." });
const stream = streamText({ model, prompt: "Explain quantum computing." });

agentId defaults to "vercel-ai".

What gets emitted

The middleware implements both wrapGenerate and wrapStream.

Identifier	axonpush eventType	Payload
llm.start	agent.start	model, prompt_tokens (non-streaming) or streaming: true
llm.token	agent.llm.token	token, one per text-delta chunk
llm.end	agent.end	model, usage, finish_reason

On the streaming path the middleware pipes the model’s stream through a TransformStream, emitting a token event per delta and reading usage and finishReason off the terminal finish chunk. Every chunk is passed through untouched, so your consumer sees exactly what it would without the middleware.

prompt_tokens on a non-streaming llm.start is the length of the prompt array, not a token count, it is captured before the call, when no token count exists yet. The real counts arrive on llm.end in usage.

Every event carries metadata.framework: "vercel-ai".