better-playwright-mcp
Documentation
better-playwright-mcp3
A high-performance Playwright MCP (Model Context Protocol) server with intelligent DOM compression and content search capabilities for browser automation.
Features
- 🎭 Full Playwright browser automation via MCP
- 🏗️ Client-server architecture with HTTP API
- 📍 Ref-based element identification system (`[ref=e1]`, `[ref=e2]`, etc.)
- 🔍 Powerful regex-based content search using ripgrep
- 💾 Persistent browser profiles with Chrome
- 🚀 91%+ DOM compression with intelligent list folding
- 📄 Semantic HTML snapshots using Playwright's internal APIs
- ⚡ High-performance search with safety limits
Installation
Global Installation (for CLI usage)
npm install -g better-playwright-mcp3Local Installation (for SDK usage)
npm install better-playwright-mcp3Usage
As a JavaScript/TypeScript SDK
Prerequisites:
1. First, start the HTTP server:
npx better-playwright-mcp3@latest server2. Then use the SDK in your code:
import { PlaywrightClient } from 'better-playwright-mcp3';
async function automateWebPage() {
// Connect to the HTTP server (must be running)
const client = new PlaywrightClient('http://localhost:3102');
// Create a page
const { pageId, success } = await client.createPage(
'my-page', // page name
'Test page', // description
'https://example.com' // URL
);
// Get page structure with intelligent folding
const outline = await client.getOutline(pageId);
console.log(outline);
// Returns compressed outline (~90% reduction) with list folding
// Search for specific content (regex by default)
const searchResult = await client.searchSnapshot(pageId, 'Example', { ignoreCase: true });
console.log(searchResult);
// Search with regular expressions (default behavior)
const prices = await client.searchSnapshot(pageId, '\\$[0-9]+\\.\\d{2}', { lineLimit: 10 });
// Search multiple patterns (OR)
const links = await client.searchSnapshot(pageId, 'link|button|input', { ignoreCase: true });
// Interact with the page using ref identifiers
await client.browserClick(pageId, 'e3'); // Click element
await client.browserType(pageId, 'e4', 'Hello World'); // Type text
await client.browserHover(pageId, 'e2'); // Hover over element
// Navigation
await client.browserNavigate(pageId, 'https://google.com');
await client.browserNavigateBack(pageId);
await client.browserNavigateForward(pageId);
// Scrolling
await client.scrollToBottom(pageId);
await client.scrollToTop(pageId);
// Waiting
await client.waitForTimeout(pageId, 2000); // Wait 2 seconds
await client.waitForSelector(pageId, 'body');
// Take screenshots
const screenshot = await client.screenshot(pageId, true); // Full page
// Clean up
await client.closePage(pageId);
}Available Methods:
- Page Management: `createPage`, `closePage`, `listPages`
- Navigation: `browserNavigate`, `browserNavigateBack`, `browserNavigateForward`
- Interaction: `browserClick`, `browserType`, `browserHover`, `browserSelectOption`, `fill`
- Advanced Actions: `browserPressKey`, `browserFileUpload`, `browserHandleDialog`
- Page Structure: `getOutline` - Get intelligently compressed page structure with list folding (NEW in v3.2.0)
- Content Search: `searchSnapshot` - Search page content with regex patterns (powered by ripgrep)
- Screenshots: `screenshot` - Capture page as image
- Scrolling: `scrollToBottom`, `scrollToTop`
- Waiting: `waitForTimeout`, `waitForSelector`
MCP Server Mode
The MCP server requires an HTTP server to be running. You need to start both:
Step 1: Start the HTTP server
npx better-playwright-mcp3@latest serverStep 2: In another terminal, start the MCP server
npx better-playwright-mcp3@latestThe MCP server will:
1. Start listening on stdio for MCP protocol messages
2. Connect to the HTTP server on port 3102
3. Route browser automation commands through the HTTP server
Standalone HTTP Server Mode
You can run the HTTP server independently:
npx better-playwright-mcp3@latest serverOptions:
- `-p, --port ` - Server port (default: 3102)
- `--host ` - Server host (default: localhost)
- `--headless` - Run browser in headless mode
- `--chromium` - Use Chromium instead of Chrome
- `--no-user-profile` - Do not use persistent user profile
- `--user-data-dir ` - User data directory
MCP Tools
When used with AI assistants, the following tools are available:
Page Management
- `createPage` - Create a new browser page with name and description
- `closePage` - Close a specific page
- `listPages` - List all managed pages with titles and URLs
Browser Actions
- `browserClick` - Click an element using its ref identifier
- `browserType` - Type text into an element
- `browserHover` - Hover over an element
- `browserSelectOption` - Select options in a dropdown
- `browserPressKey` - Press keyboard keys
- `browserFileUpload` - Upload files to file input
- `browserHandleDialog` - Handle browser dialogs (alert, confirm, prompt)
- `browserNavigate` - Navigate to a URL
- `browserNavigateBack` - Go back to previous page
- `browserNavigateForward` - Go forward to next page
- `scrollToBottom` - Scroll to bottom of page/element
- `scrollToTop` - Scroll to top of page/element
- `waitForTimeout` - Wait for specified milliseconds
- `waitForSelector` - Wait for element to appear
Content Search & Screenshots
- `searchSnapshot` - Search page content using regex patterns (powered by ripgrep)
- `screenshot` - Take a screenshot (PNG/JPEG)
Architecture
Intelligent DOM Compression (NEW in v3.2.0)
The outline generation uses a three-step compression algorithm:
1. Unwrap - Remove meaningless generic wrapper nodes
2. Text Truncation - Limit text content to 50 characters
3. List Folding - Detect and compress repetitive patterns using SimHash
Original DOM (5000+ lines)
↓
[Remove empty wrappers]
↓
[Detect similar patterns]
↓
Compressed Outline ( MCP Server HTTP Server Browser
|
v
ripgrep engineKey Design Principles
- Minimal Token Usage: Intelligent compression reduces DOM by ~91%
- On-Demand Search: Content retrieved via regex patterns when needed
- Performance: Uses ripgrep for 10x+ faster searching
- Safety: Automatic result limiting to prevent context overflow
Ref-Based Element System
Elements in snapshots are identified using ref attributes (e.g., `[ref=e1]`, `[ref=e2]`). This system:
- Provides stable identifiers for elements
- Works with Playwright's internal `aria-ref` selectors
- Enables precise element targeting across page changes
Example snapshot:
- generic [ref=e2]:
- heading "Example Domain" [level=1] [ref=e3]
- paragraph [ref=e4]: This domain is for use in illustrative examples
- link "More information..." [ref=e5] [cursor=pointer]Examples
Creating and Navigating Pages
// Create a page
const { pageId, success } = await client.createPage(
'shopping',
'Amazon shopping page',
'https://amazon.com'
);
// Navigate to another URL
await client.browserNavigate(pageId, 'https://google.com');
// Go back/forward
await client.browserNavigateBack(pageId);
await client.browserNavigateForward(pageId);Getting Page Structure (Enhanced in v3.2.0)
// Get intelligently compressed page outline
const outline = await client.getOutline(pageId);
console.log(outline);
// Example output showing list folding:
// Page Outline (473/5257 lines):
// - banner [ref=e1]
// - navigation [ref=e2]
// - list "Products" [ref=e3]
// - listitem "Product 1" [ref=e4]
// - listitem (... and 47 more similar) [refs: e5, e6, ...]
//
// Compression: 91% reduction while preserving all refsSearching Content
// Search for text (case insensitive)
const results = await client.searchSnapshot(pageId, 'product', { ignoreCase: true });
// Search with regular expression (default behavior)
const emails = await client.searchSnapshot(pageId, '[a-zA-Z0-9]+@[a-zA-Z0-9]+\\.[a-z]+');
// Search multiple patterns (OR)
const buttons = await client.searchSnapshot(pageId, 'button|submit|click', { ignoreCase: true });
// Search for prices with dollar sign
const prices = await client.searchSnapshot(pageId, '\\$\\d+\\.\\d{2}');
// Limit number of result lines
const firstTen = await client.searchSnapshot(pageId, 'item', { lineLimit: 10 });Search Options:
- `pattern` (required) - Regex pattern to search for
- `ignoreCase` (optional) - Case insensitive search (default: false)
- `lineLimit` (optional) - Maximum lines to return (default: 100, max: 100)
Response Format:
- `result` - Matched text content
- `matchCount` - Total number of matches found
- `truncated` - Whether results were truncated due to line limit
Interacting with Elements
// Click on element using its ref identifier
await client.browserClick(pageId, 'e3');
// Type text into input field
await client.browserType(pageId, 'e4', 'search query');
// Hover over element
await client.browserHover(pageId, 'e2');
// Press keyboard key
await client.browserPressKey(pageId, 'Enter');Scrolling and Waiting
// Scroll page
await client.scrollToBottom(pageId);
await client.scrollToTop(pageId);
// Wait operations
await client.waitForTimeout(pageId, 2000); // Wait 2 seconds
await client.waitForSelector(pageId, '#my-element');Best Practices for AI Assistants
Recommended Workflow: Outline First, Then Precise Actions
When using this library with AI assistants, follow this optimized workflow for maximum efficiency:
1. Start with Page Outline (Always First Step)
// Always begin by getting the compressed page structure
const outline = await client.getOutline(pageId);
// Returns intelligently compressed view with ~91% reductionThe outline provides:
- Complete page structure with intelligent list folding
- First element of each pattern preserved as sample
- All ref identifiers for precise element targeting
- Clear indication of repetitive patterns (e.g., "... and 47 more similar")
2. Use Outline to Guide Precise Searches
// Based on outline understanding, perform targeted searches
const searchResults = await client.searchSnapshot(pageId, 'specific term', {
ignoreCase: true,
lineLimit: 10
});
// Now you know exactly what to search for and where it might be3. Take Actions with Verified Ref IDs
// Use ref IDs discovered from outline or grep, not guesswork
await client.browserClick(pageId, 'e42'); // Ref ID confirmed from outlineWhy This Approach?
Token Efficiency: Compressed outline (typically = 18.0.0
- TypeScript
- Chrome or Chromium browser
Building from Source
# Clone the repository
git clone https://github.com/yourusername/better-playwright-mcp.git
cd better-playwright-mcp
# Install dependencies
npm install
# Build the project
npm run build
# Run in development mode
npm run devProject Structure
better-playwright-mcp3/
├── src/
│ ├── index.ts # Main export file
│ ├── mcp-server.ts # MCP server implementation
│ ├── client/
│ │ └── playwright-client.ts # HTTP client for browser automation
│ ├── server/
│ │ └── playwright-server.ts # HTTP server controlling browsers
│ └── utils/
│ ├── smart-outline-simple.ts # Intelligent outline generation
│ ├── list-detector.ts # Pattern detection using SimHash
│ ├── dom-simhash.ts # SimHash implementation
│ └── remove-useless-wrappers.ts # DOM cleanup
├── bin/
│ └── cli.js # CLI entry point
├── docs/
│ └── architecture.md # Detailed architecture documentation
├── package.json
├── tsconfig.json
└── README.mdTroubleshooting
Common Issues
1. Port already in use
2. Browser not launching
3. Element not found
4. Search returns too many results
Debug Mode
Enable detailed logging:
DEBUG=* npx better-playwright-mcp3Contributing
Contributions are welcome! Please feel free to submit a Pull Request.
License
MIT
Frequently asked questions
What is better-playwright-mcp?
better-playwright-mcp is a Model Context Protocol (MCP) server listed in the TrackMCP directory.
How do I install better-playwright-mcp?
Open the GitHub repository and follow its README. Most MCP servers are added to your client's MCP config, then called by your agent.
Is better-playwright-mcp open source?
Yes — it is hosted on GitHub at https://github.com/livoras/better-playwright-mcp and has 6 stars.
Related MCP tools
This sample app uses the Vercel MCP Adapter that allows you to drop in an MCP server on a group of routes in any Next.js project.
The all-in-one Desktop & Docker AI application with built-in RAG, AI agents, No-code agent builder, MCP compatibility, and more.
An AI-powered task-management system you can drop into Cursor, Lovable, Windsurf, Roo, and others. Built for the Model Context Protocol to enhance AI capabiliti
Fay是一个帮助数字人(2.5d、3d、移动、pc、网页)或大语言模型(openai兼容、deepseek)连通业务系统的mcp框架。 JavaScript-based implementation. Trusted by 12000+ developers.
CTTF: MCP integration between Cursor and Figma, allowing Cursor Agentic AI to communicate with Figma for reading designs and modifying them programmatically.
🔥 Official Firecrawl MCP Server - Adds powerful web scraping and search to Cursor, Claude and any other LLM clients. JavaScript-based implementation.
Run your own MCP server? See who uses it and what to fix.
Measure it with TrackMCP