ai-gateway-provider routes Vercel AI SDK models through Cloudflare's AI Gateway universal endpoint. You wrap pre-built @ai-sdk/* model instances and the provider captures their outbound requests, re-dispatches them through the gateway (binding or REST), and applies caching, retries, logging, and cross-vendor server-side fallback.

It shares its provider registry and cf-aig-* header building with the workers-ai-provider AI Gateway delegate, so the two stay in lockstep.

npm install ai-gateway-provider ai

Install only the @ai-sdk/* providers you actually wrap — they're optional peers:

npm install @ai-sdk/openai      # also covers the OpenAI-compatible long tail
npm install @ai-sdk/anthropic

Wrap a single pre-built model and route it through your gateway. Inside a Worker, use the env.AI.gateway(id) binding; the rest of your AI SDK code is unchanged.

import { createAiGateway } from "ai-gateway-provider";
import { createOpenAI } from "ai-gateway-provider/providers/openai";
import { generateText } from "ai";

const aigateway = createAiGateway({ binding: env.AI.gateway("my-gateway") });
const openai = createOpenAI({ apiKey: env.OPENAI_API_KEY });

const { text } = await generateText({
	model: aigateway([openai("gpt-5")]),
	prompt: "Write a haiku about Cloudflare Workers.",
});

Outside a Worker (Node, scripts, CI), use the REST path with your account id, gateway id, and — if the gateway has authentication enabled — its token:

const aigateway = createAiGateway({
	accountId: process.env.CLOUDFLARE_ACCOUNT_ID!,
	gateway: "my-gateway",
	apiKey: process.env.CLOUDFLARE_AI_GATEWAY_TOKEN, // only if the gateway is authenticated
});

Streaming, structured output, and tools all work — they're just standard AI SDK calls on the wrapped model:

import { generateObject, streamText } from "ai";
import { z } from "zod";

const model = aigateway([openai("gpt-5")]);

// Streaming
const result = streamText({ model, prompt: "Tell me a story." });
for await (const chunk of result.textStream) process.stdout.write(chunk);

// Structured output
const { object } = await generateObject({
	model,
	prompt: "Describe France.",
	schema: z.object({ capital: z.string(), populationMillions: z.number() }),
});

Both reach the same AI Gateway. Pick based on how you construct models:

Both support resumable streaming (coming soon) on the binding/run path (cf-aig-run-id): the delegate enables it by default, while ai-gateway-provider exposes it as an opt-in resume setting (see below).

Pass multiple models; the gateway tries them in order and the first success wins (selected via the cf-aig-step response header):

import { createAiGateway } from "ai-gateway-provider";
import { createOpenAI } from "ai-gateway-provider/providers/openai";
import { createAnthropic } from "ai-gateway-provider/providers/anthropic";

const aigateway = createAiGateway({ binding: env.AI.gateway("my-gateway") });
const openai = createOpenAI({ apiKey: OPENAI_API_KEY });
const anthropic = createAnthropic({ apiKey: ANTHROPIC_API_KEY });

const model = aigateway([
	openai.chat("gpt-5"),
	anthropic("claude-sonnet-4-5"), // cross-vendor fallback
]);

options on createAiGateway map to the current cf-aig-* request headers:

The deprecated cf-cache-ttl / cf-skip-cache names are no longer emitted.

Set them once on the provider and they apply to every request:

const aigateway = createAiGateway({
	binding: env.AI.gateway("my-gateway"),
	options: {
		cacheTtl: 3600, // cache responses for an hour
		metadata: { team: "search", userId: 12345 }, // spend attribution in the dashboard
	},
});

🚧 Coming soon. Resumable streaming is not generally available yet — the AI Gateway resume backend is still rolling out. The resume option below is in place so you can adopt it early, but treat resume as experimental until the rollout completes.

On the binding (run) path, pass resume with the full env.AI binding and the gateway id to recover transient mid-stream drops via the gateway resume endpoint:

const aigateway = createAiGateway({
	binding: env.AI.gateway("my-gateway"),
	resume: { binding: env.AI, gateway: "my-gateway" }, // onResumeExpired?: "error" | "accept-partial"
});

It uses the same resumable-stream engine as the workers-ai-provider delegate. Resume is a no-op when no cf-aig-run-id is present — e.g. the REST/API-key transport or non-streaming calls.

See gateway routing for the full transport and fallback model.