@useautumn/gateway weekly downloads statistics

@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

Requires Node.js 22+ and ESM imports. The AI SDK adapter targets AI SDK 7 and V4 language models; AI SDK 6 applications should stay on the previous gateway release.

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

The wrapped model is a drop-in LanguageModelV4. Usage is read per model call from the response (or the stream's finish chunk), including each step in a tool loop, and reported as / (e.g. anthropic/claude-sonnet-4-5). Tracking does not use AI SDK 7's aggregated top-level result.usage.

OpenAI, Anthropic, and Google provider names are normalized automatically: plain names and dotted SDK namespaces such as openai.responses, anthropic.messages, and google.generative-ai use their models.dev provider keys. Vercel's gateway namespace maps to vercel. Model IDs retain all dots and slashes. Other provider names pass through unchanged; use providerId for a custom mapping. Explicit overrides are always used verbatim.

OpenRouter

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/ 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

Description from the npm listing.