Cloudflare Playwright

You can run a Playwright browser controller as an MCP server and connect it to AI tools and copilots. This setup lets you navigate web pages, type, click, and capture screenshots through a structured set of tools exposed by the MCP server.

How to use

This MCP server enables you to control a browser through a defined set of tools. Use an MCP client or platform that supports SSE-based MCP servers to connect to the server URL and start issuing tasks like navigating to a page, typing into fields, clicking elements, and taking screenshots. After connection, you can instruct the model with simple actions, such as “Go to a page” or “Take a screenshot,” and the model will perform the corresponding browser actions through the available tools.

How to install

Prerequisites: Node.js and npm or a compatible runtime for MCP setup. You will deploy and run the MCP server as described here.

# Install dependencies for the project
npm ci

# Publish or deploy the MCP server to Cloudflare Workers
npx wrangler deploy

Configuration and usage notes

You can connect to the Cloudflare Playwright MCP server from an MCP client or platform that supports SSE endpoints. Use the following URL format for the remote server: https://[my-mcp-url].workers.dev/sse. For local testing with Claude Desktop via a proxy, you can configure a local MCP proxy to forward to the remote SSE URL.

Security and access: Ensure you deploy to an environment where the SSE endpoint is reachable from your client. If you are using Claude Desktop, you may run a local proxy (mcp-remote) to forward the remote MCP server to your local environment.

Troubleshooting and tips

If you encounter connectivity issues, verify that the SSE URL is reachable from the client network, and confirm that the deployment to the Cloudflare Workers endpoint completed successfully. If the session fails to connect, check the network configuration and ensure that your MCP client is configured to use the correct SSE URL.

Examples and workflow notes

A typical session involves opening a target page, performing actions such as typing or clicking, and taking screenshots to verify results. Use simple, single-action instructions first to achieve reliable tool execution, then build more complex sequences as needed.

Available tools

browser_navigate

Navigate the browser to a specified URL and wait until the page load completes.

browser_type

Type text into an input field or editable element, with optional delay between keystrokes.

browser_clickt

Click a targeted element, such as a button or checkbox, to trigger UI interactions.

browser_take_screenshot

Capture a screenshot of the current browser view and return it in PNG format.