# Cloudflare Computer — Headless Browser Environment for AI Agents

> A cloud-native headless browser runtime by Cloudflare that gives AI agents full computer-use capabilities including browsing, clicking, typing, and taking screenshots.

## Install

Save in your project root:

# Cloudflare Computer — Headless Browser Environment for AI Agents

## Quick Use
```bash
npm install @cloudflare/computer
# In your agent code:
# import { Computer } from "@cloudflare/computer"
# const computer = new Computer()
# await computer.navigate("https://example.com")
# const screenshot = await computer.screenshot()
```

## Introduction
Cloudflare Computer is an open-source TypeScript library that provisions a full headless browser environment for AI agents. It lets agents navigate websites, fill forms, click buttons, read page content, and take screenshots — all through a simple programmatic API. Built on Cloudflare's edge infrastructure, it provides fast, isolated browser sessions that agents can use as their eyes and hands on the web.

## What Cloudflare Computer Does
- Provisions isolated headless browser sessions for each agent request
- Supports navigation, clicking, typing, scrolling, and screenshot capture
- Returns structured page content (text, links, form fields) alongside visual screenshots
- Handles authentication flows including cookie persistence across page loads
- Integrates with MCP and standard agent tool-use protocols

## Architecture Overview
The library wraps a Chromium-based headless browser running on Cloudflare Workers. Each session gets its own isolated browser context with a fresh profile, preventing cross-session data leakage. The TypeScript SDK communicates with the browser backend over a WebSocket connection, streaming commands and receiving results with low latency. Sessions are ephemeral by default but can persist state across multiple interactions when configured.

## Self-Hosting & Configuration
- Install the npm package and import it into your TypeScript or JavaScript agent
- Configure the Cloudflare account credentials or run locally with a bundled Chromium instance
- Set viewport dimensions, user-agent strings, and timeout values via the constructor
- Enable cookie persistence for multi-step workflows that require login state
- Use environment variables to control concurrency limits and session lifetimes

## Key Features
- Full browser automation: navigate, click, type, scroll, screenshot, and extract content
- Session isolation ensures agents never leak state between tasks
- Works with any LLM agent framework that supports tool-use or function calling
- Built-in content extraction returns clean text alongside raw HTML
- Low-latency edge execution via Cloudflare's global network

## Comparison with Similar Tools
- **Browser Use** — Python-based browser automation for agents; Cloudflare Computer is TypeScript-native and edge-hosted
- **Stagehand** — focuses on natural-language browser instructions; Cloudflare Computer offers lower-level programmatic control
- **Playwright** — general browser testing framework; Cloudflare Computer is purpose-built for agent computer-use
- **E2B** — sandboxed code execution; Cloudflare Computer specializes in browser interactions

## FAQ
**Q: Do I need a Cloudflare account?**
A: For edge-hosted sessions, yes. For local development, the SDK can run against a local Chromium.

**Q: Can agents interact with JavaScript-heavy single-page apps?**
A: Yes. The headless browser fully renders JavaScript before returning page content.

**Q: Is there a session time limit?**
A: Default sessions expire after 5 minutes of inactivity but the timeout is configurable.

**Q: Which agent frameworks are supported?**
A: Any framework that can call TypeScript functions. Adapters exist for MCP, LangChain, and Vercel AI SDK.

## Sources
- https://github.com/cloudflare/computer

---
Source: https://tokrepo.com/en/workflows/asset-3243f2b9
Author: AI Open Source