@useautumn/gateway

Autumn adapters for AI SDKs and gateways. Wrap your model or client once and every LLM call's token usage is tracked against the customer's AI credit balance — no manual trackTokens calls.

Autumn converts the token counts to a dollar cost server-side using live models.dev rates plus your configured markup, so pricing changes never require a client redeploy.

Install

npm install @useautumn/gateway
# or
bun add @useautumn/gateway

Authentication: pass an autumn client (from autumn-js), or set AUTUMN_API_KEY (or AUTUMN_SECRET_KEY) in the environment and omit it.

Vercel AI SDK

import { openai } from "@ai-sdk/openai";
import { generateText } from "ai";
import { withAutumn } from "@useautumn/gateway/ai-sdk";

const model = withAutumn({
  model: openai("gpt-4o"),
  customerId: "user_123",
});

const { text } = await generateText({ model, prompt: "Hello!" });
// usage tracked automatically — streaming too

The wrapped model is a drop-in LanguageModelV3. Usage is read from the response (or the stream's finish chunk) and reported as <provider>/<model> (e.g. anthropic/claude-sonnet-4-5). If your provider's name doesn't match its models.dev key, override it with providerId.

OpenRouter

import { OpenRouter } from "@openrouter/sdk";
import { withAutumn, trackingSettled } from "@useautumn/gateway/openrouter";

const client = withAutumn({
  openRouter: new OpenRouter({ apiKey: process.env.OPENROUTER_API_KEY }),
  customerId: "user_123",
});

const result = await client.chat.send({
  model: "openai/gpt-4o",
  messages: [{ role: "user", content: "Hello!" }],
});

The wrapper forces OpenRouter usage accounting on every request, tracks chat.send (streaming and non-streaming) and the responses API (including @openrouter/agent's callModel), and reports models as openrouter/<slug> using the response's resolved model — router aliases like openrouter/auto bill against the model that actually served the request. OpenRouter's own reported charge is attached to each event as openrouter_cost.

Streaming usage is tracked in the background; call await trackingSettled(client) before reading balances that must reflect calls just made. For consumption patterns the wrapper doesn't cover, trackOpenRouterUsage({ usage, model, customerId }) accepts both SDK camelCase and raw snake_case usage objects.

Options

Both adapters accept the shared tracking options:

| Option | Required | Description | |--------|----------|-------------| | customerId | Yes | Autumn customer ID to attribute usage to | | autumn | No | autumn-js client; falls back to AUTUMN_API_KEY env | | featureId | No | Target AI credit system feature (auto-detected if you have one) | | entityId | No | Entity for entity-scoped balances | | properties | No | Extra properties attached to each usage event |

Tracking failures are caught and logged — they never break your AI responses.

Docs

• Vercel AI SDK guide
• OpenRouter guide
• trackTokens API reference