Overview
Agent Starter is Cloudflare’s official template for building AI chat agents with the Agents SDK. It gives you a working chat app — streaming responses, tools, scheduling, and image understanding — running on Workers AI, so you can start customizing an agent in minutes without a third-party model API key.
What’s Included
- Streaming AI chat powered by Workers AI through
AIChatAgent - Three tool patterns: server-side tools that run automatically, client-side tools answered by the browser, and human-in-the-loop tools that ask for approval first
- Task scheduling: one-time, delayed, and recurring (cron) tasks
- Vision: drag, paste, or attach images for vision-capable models
- Reasoning display that streams model thinking and collapses when done
- Debug mode to inspect the raw message JSON
- Real-time WebSocket connection with automatic reconnection and persisted messages
- UI built with Kumo, Cloudflare’s design system, with light and dark themes
Getting Started
npx create-cloudflare@latest --template cloudflare/agents-starter
cd agents-starter
npm install
npm run dev
Then open http://localhost:5173. Because Workers AI has no local simulator, local development uses a remote session — run wrangler login once or set a CLOUDFLARE_API_TOKEN. No OpenAI or Anthropic key is required.
Try prompts such as “What’s the weather in Paris?”, “Calculate 5000 * 3” (which asks for approval), or “Remind me in 5 minutes to take a break”.
Project Structure
| File | Role |
|---|---|
src/server.ts |
The chat agent, its tools, and scheduling |
src/app.tsx |
Chat UI built with Kumo components |
src/client.tsx |
React entry point |
src/styles.css |
Tailwind and Kumo styles |
Making It Your Own
Rename the project in package.json and wrangler.jsonc (the Worker name becomes your workers.dev URL), add your own tools in server.ts, and deploy with Wrangler.
License
MIT License.