If you are building an AI native application, this package is the fastest way to monetize it. It wraps the official @ai-sdk functions, intercepts token consumption, and reports exactly what model and how many tokens were used—without breaking your streams.

Key Features

• Zero Friction: Drop-in replacements for streamText and generateText.
• Smart Backend: We transmit the raw usage (promptTokens, completionTokens, modelId), and Revstack Cloud calculates the exact credit deduction based on your configured margins and model pricing.
• Non-Blocking: Usage tracking occurs transparently and never delays the user's stream.

Installation

npm install @revstackhq/ai ai

Quick Start

1. Configure Revstack Once (IoC)

Create a pre-configured instance of the AI wrapper in your app. Pass a trackUsage callback so @revstackhq/ai can report usage without knowing about your framework.

// lib/revstack.ts
import { trackUsage } from "@revstackhq/next/server";
import { createRevstackAI } from "@revstackhq/ai";

export const revstack = createRevstackAI(
  { secretKey: process.env.REVSTACK_SECRET_KEY! },
  async (key, usage, config) => {
    // This fires every time a stream or generation completes
    await trackUsage(key, usage, config);
  }
);

2. Replace streamText and generateText

Import your pre-configured instance instead of the base Vercel functions. You now only need to provide an entitlementKey.

// app/api/chat/route.ts
import { revstack } from "@/lib/revstack";
import { openai } from "@ai-sdk/openai";

export async function POST(req: Request) {
  const { messages } = await req.json();

  const result = await revstack.streamText({
    model: openai("gpt-4o"),
    messages,
    entitlementKey: "ai_tokens", // Triggers metering automatically
    // Your original onFinish still fires!
    async onFinish(event) {
      console.log("Stream completed locally, and usage was already tracked.");
    },
  });

  return result.toDataStreamResponse();
}

How It Works

When a user streams a generation:

1. The stream begins sending chunks to the client instantly.
2. Vercel's internal onFinish event fires when the stream ends.
3. Your pre-configured revstack.streamText wrapper intercepts this event, extracting the model.id and exact token split.
4. It calls your injected trackUsage callback, transmitting the raw payload (e.g. promptTokens: 10, completionTokens: 50) to your backend.
5. If you provided an onFinish handler, it runs sequentially afterward.

This offloads the complex math of "how many credits should GPT-4o cost vs Claude 3.5 Sonnet" entirely to your Revstack product configuration.