MCP 服务器

Cloudflare Playwright MCP

An MCP server that enables AI agents to automate browser interactions using Playwright and Cloudflare Workers, supporting tasks like navigation, clicking, typing, and screenshots.

README

Cloudflare Playwright MCP

Overview

This project leverages Playwright for automated browser testing and integrates with Cloudflare Workers, Browser Rendering and @cloudflare/playwright for deployment.

Build and Deploy

Follow these steps to set up and deploy the project:

Install dependencies:

npm ci

Build:

cd cloudflare
npm run build

Deploy to Cloudflare Workers:

cd cloudflare/example
npm ci
npx wrangler deploy

Use with Cloudflare AI Playground

alt text

Cloudflare playground AI is a great way to test MCP servers using LLM models available in Workers AI.

Navigate to https://playground.ai.cloudflare.com/
Ensure model is set to llama-3.3-70b-instruct-fp8-fast
In MCP Servers, set URL to https://[my-mcp-url].workers.dev/sse
Click Connect
Status should update to Connected and it should list 14 available tools

You can now start to interact with the model, and it will run necessary tools to accomplish what was requested.

[!TIP] For best results, give simple instructions consisting of one single action, e. g., "Create a new todo entry", "Go to cloudflare site", "Take a screenshot"

Example of a conversation:

Role	Message
User	Go to demo.playwright.dev/todomvc
Assistant	runs browser_navigate tool<br> The page has been navigated to https://demo.playwright.dev/todomvc/. The page title is "React • TodoMVC". The page snapshot shows the content of the page, including a heading, a textbox, and several links.
User	Create some todo entry
Assistant	runs browser_type tool<br> A new todo entry "Buy groceries" has been created. The page snapshot shows the updated content of the page, including the new todo entry.
User	Nice. Now create a todo in parrot style
Assistant	runs browser_type tool<br> A new todo entry "Polly wants a cracker" has been created in parrot style. The page snapshot shows the updated content of the page, including the new todo entry.
User	And create another todo in yoda style
Assistant	runs browser_type tool<br> A new todo entry "A walk, I must take" has been created in Yoda style. The page snapshot shows the updated content of the page, including the new todo entry.
User	Take a screenshot
Assistant	runs browser_take_screenshot tool<br> A screenshot of the current page has been taken. The screenshot is in PNG format and is 56KB in size.
User	Now click the checkbox of the parrot todo entry
Assistant	runs browser_clickt tool<br> The checkbox of the parrot todo entry "Polly wants a cracker" has been clicked. The page snapshot shows the updated content of the page, with the checkbox now checked.
User	Take a screenshot
Assistant	runs browser_take_screenshot tool<br> A screenshot of the current page has been taken. The screenshot is in PNG format and is 58KB in size.
User	Remove completed todo entries
Assistant	runs browser_clickt tool<br> The "Clear completed" button has been clicked. The page snapshot shows the updated content of the page, with the completed todo entry "Polly wants a cracker" removed.
User	Take a final screenshot
Assistant	runs browser_take_screenshot tool<br> A final screenshot of the current page has been taken. The screenshot is in PNG format and is 50KB in size.

Use with Claude Desktop

As of now, Claude Desktop only supports local MCP servers. To use playground-mcp with Claude Desktop we make use of mcp-remote, a tool that proxies remote MCP servers and exposes them locally. Use the following configuration:

Open the configuration file for Claude Desktop.
Add the following JSON snippet under the mcpServers section:

{
  "mcpServers": {
    "cloudflare-playwright-mcp": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://[my-mcp-url].workers.dev/sse"
      ]
    }
  }
}

Save the configuration file and restart Claude Desktop to apply the changes.

This setup ensures that Claude Desktop can communicate with the Cloudflare Playwright MCP server.

Here's an example of a session opening the TODO demo app, adding "buy lemons" and doing a screenshot, taking advantage of playwright-mcp tools and Browser Rendering:

alt text

Configure in VSCode

You can install the Playwright MCP server using the VS Code CLI:

# For VS Code
code --add-mcp '{"name":"cloudflare-playwright","type":"sse","url":"https://[my-mcp-url].workers.dev/sse"}'

# For VS Code Insiders
code-insiders --add-mcp '{"name":"cloudflare-playwright","type":"sse","url":"https://[my-mcp-url].workers.dev/sse"}'

After installation, the Playwright MCP server will be available for use with your GitHub Copilot agent in VS Code.

Use with Cursor

By default, when imageResponses is set to "auto", the server disables inline image responses for Cursor clients. To enable inline screenshots in Cursor, you need to override this behavior:

If you're running your own Cloudflare Worker, pass imageResponses: "allow" when creating the MCP agent:

export const PlaywrightMCP = createMcpAgent(env.BROWSER, {
  imageResponses: 'allow',
} as any);

Tool Modes

The tools are available in two modes:

Snapshot Mode (default): Uses accessibility snapshots for better performance and reliability
Vision Mode: Uses screenshots for visual-based interactions

Vision Mode works best with the computer use models that are able to interact with elements using X Y coordinate space, based on the provided screenshot.

<details> <summary>Interactions</summary>