Workers AI provider for the Vercel AI SDK. Run Workers AI models, and route any catalog model through AI Gateway with unified billing, server-side fallback, and resumable streaming (coming soon).

• Package README (quickstart): packages/workers-ai-provider
• Concepts: gateway routing · resume (coming soon) · binding vs REST

🚧 Resumable streaming is coming soon. It's not generally available yet — the AI Gateway resume backend is still rolling out. The resume / onResumeExpired options below are in place so you can adopt them early, but treat resume as experimental until the rollout completes.

import { createWorkersAI } from "workers-ai-provider";
import { generateText } from "ai";

const workersai = createWorkersAI({ binding: env.AI });

const { text } = await generateText({
	model: workersai("@cf/zai-org/glm-5.2"),
	prompt: "Write a haiku about Cloudflare Workers.",
});

Supports chat (tool calling, reasoning), image generation, embeddings, transcription, text-to-speech, and reranking. Works with a binding or with accountId + apiKey (REST).

Wrap any @ai-sdk/* provider so its requests are routed through your gateway (host-stripped to the gateway-native endpoint, with cf-aig-* headers applied). Useful for provider-native or non-chat providers the slug delegate can't auto-wire, or when you want full control over the model instance.

import { createGatewayProvider } from "workers-ai-provider";
import { createOpenAI } from "@ai-sdk/openai";
import { generateText } from "ai";

const openai = createGatewayProvider(createOpenAI, {
	binding: env.AI,
	gateway: "my-gateway",
	byok: true, // forward the provider key; omit to use unified billing / a stored key
	apiKey: env.OPENAI_API_KEY,
});

const { text } = await generateText({ model: openai("gpt-5"), prompt: "Hello" });

The provider id is detected from the request URL. Outside a Worker, the lower-level createGatewayFetch from workers-ai-provider/gateway does the same with explicit credentials.

Configure createWorkersAI with provider plugins, then pass a vendor/model slug (instead of a @cf/... id) to route it through AI Gateway with capability-driven transport selection (run path vs gateway path), resume, caching, and fallback.

import { createWorkersAI } from "workers-ai-provider";
import { openai } from "workers-ai-provider/openai";
import { anthropic } from "workers-ai-provider/anthropic";

const gatewayAi = createWorkersAI({
	binding: env.AI,
	gateway: { id: "my-gateway" },
	providers: [openai, anthropic],
});

// Run path + resume (default for unified-catalog providers):
const model = gatewayAi("openai/gpt-5");

// Cross-vendor server-side fallback (one gateway run, cf-aig-step picks winner):
const resilient = gatewayAi("openai/gpt-5", {
	fallback: { mode: "server", models: ["anthropic/claude-sonnet-4-5"] },
});

Provider plugins are imported from sub-paths so the @ai-sdk/* packages stay optional peers: workers-ai-provider/openai, /anthropic, /google. One plugin covers a whole wire format — the openai plugin serves the entire OpenAI-compatible long tail. A @cf/... id on the same provider still routes to Workers AI directly, so one provider instance handles both.

See gateway routing for how transport, fallback, and cacheTtl interact with transport selection.

A single fetch handler that streams a resilient, resumable response. Workers AI models and catalog slugs share the same provider instance.

import { createWorkersAI } from "workers-ai-provider";
import { openai } from "workers-ai-provider/openai";
import { anthropic } from "workers-ai-provider/anthropic";
import { streamText } from "ai";

export default {
	async fetch(req: Request, env: { AI: Ai }) {
		const workersai = createWorkersAI({
			binding: env.AI,
			providers: [openai, anthropic],
			gateway: { id: "my-gateway" },
		});

		const { prompt } = (await req.json()) as { prompt: string };

		const result = streamText({
			// Run path + resume by default; client-side fallback keeps resume per leg.
			model: workersai("openai/gpt-5", {
				fallback: { mode: "client", models: ["anthropic/claude-sonnet-4-5"] },
				onResumeExpired: "accept-partial",
			}),
			prompt,
		});

		return result.toTextStreamResponse();
	},
};

• createWorkersAI (catalog slug routing is enabled by passing providers), createGatewayProvider
• GATEWAY_PROVIDERS, findProviderBySlug, detectProviderByUrl, wireableProviders (the shared registry)
• createResumableStream, GatewayDelegateError, WorkersAIGatewayError, WorkersAIFallbackError, createClientFallbackModel
• DelegateCallOptions, ProviderPlugin, Transport, WireFormat (types)
• Provider plugins from workers-ai-provider/{openai,anthropic,google}

The shared core is bundled into each consumer, so shared error classes are distinct per bundle — cross-package instanceof would fail. Match shared errors by .name / a kind discriminant rather than instanceof. Catching the error class exported by this package (single bundle) works fine.