Agents Starter is Cloudflare’s template for building AI chat agents on Workers, powered by their Agents SDK. It ships a working chat agent with streaming responses, tool use, scheduling, and vision — all running on Workers AI with no API key required. This is my fork, kept for experimenting with agent patterns on Cloudflare’s edge.
The concepts
What makes this template useful is that it demonstrates three distinct tool patterns in one place: server-side tools that auto-execute, client-side tools where the browser provides the answer (like timezone detection), and human-in-the-loop tools that ask for approval before running. On top of that it shows scheduling — one-time, delayed, and recurring cron tasks — plus a reasoning display that streams the model’s thinking, image input for vision-capable models, and WebSocket-based real-time messaging with persistence. It is a compact tour of how the Agents SDK wires an LLM into a stateful, tool-using service.
How to use it
The most impactful change is editing the system prompt in server.ts to give the agent its personality and focus. From there you replace the demo tools (a fake weather tool, a basic calculator) with real API calls, and add your own tools to the tools object following whichever of the three patterns fits.
Running it
npx create-cloudflare@latest --template cloudflare/agents-starter
cd agents-starter
npm install
npm run dev
Open localhost:5173 and try prompts like “What’s the weather in Paris?” or “Remind me in 5 minutes to take a break” to exercise the different tool types.
Language: TypeScript Source code: github.com/avnit/agents-starter Upstream: github.com/cloudflare/agents