Run an open chat model from a Cloudflare Worker with Workers AI

What you get

A Cloudflare Worker that sends a chat conversation to Google's open Gemma 4 model on Workers AI and returns the reply as JSON, with no API key to manage.

Prerequisites

A Cloudflare account (Workers AI is included on the Free and Paid Workers plans; the free allocation is 10,000 Neurons per day)
Node.js and npm installed

Steps

Create a new Worker project. When prompted pick Hello World example, Worker only, TypeScript, Git Yes, deploy now No.

Shell

npm create cloudflare@latest -- hello-ai
cd hello-ai

Connect the Worker to Workers AI. Open the generated wrangler.jsonc and add the ai binding inside the existing top-level object:

jsonc

{
	"ai": {
		"binding": "AI"
	}
}

If your project uses wrangler.toml instead, add:

TOML

[ai]
binding = "AI"

Replace the contents of src/index.ts with:

TypeScript

export interface Env {
	AI: Ai;
}

export default {
	async fetch(request, env): Promise<Response> {
		const response = await env.AI.run("@cf/google/gemma-4-26b-a4b-it", {
			messages: [
				{
					role: "system",
					content: "You are a helpful assistant.",
				},
				{
					role: "user",
					content: "What is the origin of the phrase Hello, World",
				},
			],
			chat_template_kwargs: {
				enable_thinking: false,
			},
		});

		return Response.json(response);
	},
} satisfies ExportedHandler<Env>;

Run the Worker locally. Wrangler may ask you to log in, because Workers AI inference runs on Cloudflare even in local mode (and counts toward your limits).

Shell

npx wrangler dev

Shell

npx wrangler login
npx wrangler deploy

Test it

With npx wrangler dev running, open the local URL Wrangler prints (usually http://localhost:8787) in your browser. You should see a JSON response containing Gemma's answer about the origin of "Hello, World". After deploying, the same response is served at https://hello-ai.<YOUR_SUBDOMAIN>.workers.dev. Change the user message in src/index.ts to ask something else.