Vercel AI Gateway

The Vercel AI Gateway speaks the OpenAI and Anthropic wire protocols, so you reach it with an ordinary provider client pointed at a different base URL. How you capture it depends on which language you are in, and the two answers are genuinely different.

Language	How to capture gateway traffic
Python	Wrap an OpenAI or Anthropic client whose base URL is the gateway
TypeScript	Use the AI SDK middleware on a gateway() model

Python

Point the official client at the gateway and wrap it. Metergraph recognizes the public gateway host automatically.

app/llm.py

import os
import metergraph
from openai import OpenAI

gateway = metergraph.wrap(OpenAI(
    api_key=os.getenv("AI_GATEWAY_API_KEY") or os.getenv("VERCEL_OIDC_TOKEN"),
    base_url="https://ai-gateway.vercel.sh/v1",
))

gateway.chat.completions.create(
    model="anthropic/claude-sonnet-4.6",
    messages=[{"role": "user", "content": "Hello"}],
)

An Anthropic client pointed at the same host works the same way.

Detection is an exact hostname match on a parsed URL, never a substring. The host must be ai-gateway.vercel.sh. A path such as /v1 makes no difference, and a lookalike host such as ai-gateway.vercel.sh.example.com is not detected.

Custom gateway URLs

If your gateway sits behind your own domain, the host will not match. Declare it:

gateway = metergraph.wrap(client, provider="vercel")

"gateway" and "vercel-ai-gateway" are accepted as the same thing.

The vercel provider name is Python only

The TypeScript wrap() accepts "openai", "anthropic" and "google" and nothing else. There is no "vercel" provider there, and no automatic gateway host detection in the Node wrap(). Use the middleware instead, as below.

What the row looks like

On the gateway path the request model is a creator-qualified identifier such as anthropic/claude-sonnet-4.6. Metergraph takes the provider from the part before the slash, and keeps the full identifier as the model. So a call to anthropic/claude-sonnet-4.6 is recorded with provider anthropic, and sits alongside your direct Anthropic calls in the dashboard rather than in a category of its own. A model with no slash is recorded with the provider vercel-ai-gateway.

Pricing is resolved on the server from the effective-dated catalog, at read time. A model the catalog does not know stays visible and lands in the unpriced bucket instead of being dropped. See Model catalog and pricing.

Streaming behaves exactly as it does for a direct OpenAI chat completion, including the stream_options handling described under OpenAI.

The batch-result readers are not attached on the gateway path. Only the inference seams are patched.

Never let the gateway credential into capture

AI_GATEWAY_API_KEY and VERCEL_OIDC_TOKEN are credentials. They live on the client, not in the request body, so they are not part of what Metergraph captures. Keep it that way: do not copy them into a header field you then pass through the request payload, and do not put them in a tag.

TypeScript

The Node wrap() does not detect the gateway host. Capture gateway traffic through the AI SDK middleware instead, which canonicalizes the provider from the same creator/model prefix:

src/llm.ts

import * as mg from "metergraph";
import { gateway, generateText, wrapLanguageModel } from "ai";

const model = wrapLanguageModel({
  model: gateway("anthropic/claude-sonnet-4.5"),
  middleware: mg.vercelAISDKMiddleware(),
});

await mg.trace("support-answer", () =>
  generateText({ model, prompt: "Help this customer" }));

The middleware recognizes gateway, vercel and vercel-ai-gateway as gateway provider names and resolves the real provider from the model identifier’s creator prefix, so anthropic/claude-sonnet-4.5 is recorded as anthropic here too.

Full setup, including the AI SDK version options, is on Vercel AI SDK.

Do not do both

If a call already goes through the middleware, do not also wrap the underlying client. One provider request would be captured twice.

Next

• Vercel AI SDK for the middleware in full
• OpenRouter for the other supported gateway
• Model catalog and pricing for how a creator/model identifier is priced