Playwright MCP

A Model Context Protocol (MCP) server that provides browser automation capabilities using Playwright. This server enables LLMs to interact with web pages through structured accessibility snapshots, bypassing the need for screenshots or visually-tuned models.

Key Features

• Fast and lightweight: Uses Playwright's accessibility tree, not pixel-based input.
• LLM-friendly: No vision models needed, operates purely on structured data.
• Deterministic tool application: Avoids ambiguity common with screenshot-based approaches.
• Computer Use Agent (CUA): Supports high-level agent-based browser automation.

Use Cases

• Web navigation and form-filling
• Data extraction from structured content
• Automated testing driven by LLMs
• General-purpose browser interaction for agents
• Long-running browser automation with Computer Use Agent (CUA)

Example config

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": [
        "@playwright/mcp@latest"
      ]
    }
  }
}

Installation in VS Code

Install the Playwright MCP server in VS Code using one of these buttons:

Alternatively, you can install the Playwright MCP server using the VS Code CLI:

# For VS Code
code --add-mcp '{"name":"playwright","command":"npx","args":["@playwright/mcp@latest"]}'

# For VS Code Insiders
code-insiders --add-mcp '{"name":"playwright","command":"npx","args":["@playwright/mcp@latest"]}'

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

CLI Options

The Playwright MCP server supports the following command-line options:

• --browser <browser>: Browser or chrome channel to use. Possible values:
  • chrome, firefox, webkit, msedge
  • Chrome channels: chrome-beta, chrome-canary, chrome-dev
  • Edge channels: msedge-beta, msedge-canary, msedge-dev
  • Default: chrome
• --cdp-endpoint <endpoint>: CDP endpoint to connect to
• --executable-path <path>: Path to the browser executable
• --headless: Run browser in headless mode (headed by default)
• --port <port>: Port to listen on for SSE transport
• --user-data-dir <path>: Path to the user data directory
• --vision: Run server that uses screenshots (Aria snapshots are used by default)

User data directory

Playwright MCP will launch the browser with the new profile, located at

- `%USERPROFILE%\AppData\Local\ms-playwright\mcp-chrome-profile` on Windows
- `~/Library/Caches/ms-playwright/mcp-chrome-profile` on macOS
- `~/.cache/ms-playwright/mcp-chrome-profile` on Linux

All the logged in information will be stored in that profile, you can delete it between sessions if you'd like to clear the offline state.

Chrome Extensions Support

Chrome extensions are enabled by default when using a Chromium-based browser in headed mode. This allows you to access the Chrome Web Store and install extensions directly from there.

Running headless browser (Browser without GUI).

This mode is useful for background or batch operations.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": [
        "@playwright/mcp@latest",
        "--headless"
      ]
    }
  }
}

Running headed browser on Linux w/o DISPLAY

When running headed browser on system w/o display or from worker processes of the IDEs, run the MCP server from environment with the DISPLAY and pass the --port flag to enable SSE transport.

npx @playwright/mcp@latest --port 8931

And then in MCP client config, set the url to the SSE endpoint:

{
  "mcpServers": {
    "playwright": {
      "url": "http://localhost:8931/sse"
    }
  }
}

Tool Modes

The tools are available in two modes:

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

To use Vision Mode, add the --vision flag when starting the server:

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": [
        "@playwright/mcp@latest",
        "--vision"
      ]
    }
  }
}

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.

Programmatic usage with custom transports

import { createServer } from '@playwright/mcp';

// ...

const server = createServer({
  launchOptions: { headless: true }
});
transport = new SSEServerTransport("/messages", res);
server.connect(transport);

Snapshot Mode

The Playwright MCP provides a set of tools for browser automation. Here are all available tools:

• browser_navigate

  • Description: Navigate to a URL
  • Parameters:
    • url (string): The URL to navigate to

• browser_go_back

  • Description: Go back to the previous page
  • Parameters: None

• browser_go_forward

  • Description: Go forward to the next page
  • Parameters: None

• browser_click

  • Description: Perform click on a web page
  • Parameters:
    • element (string): Human-readable element description used to obtain permission to interact with the element
    • ref (string): Exact target element reference from the page snapshot

• browser_hover

  • Description: Hover over element on page
  • Parameters:
    • element (string): Human-readable element description used to obtain permission to interact with the element
    • ref (string): Exact target element reference from the page snapshot

• browser_drag

  • Description: Perform drag and drop between two elements
  • Parameters:
    • startElement (string): Human-readable source element description used to obtain permission to interact with the element
    • startRef (string): Exact source element reference from the page snapshot
    • endElement (string): Human-readable target element description used to obtain permission to interact with the element
    • endRef (string): Exact target element reference from the page snapshot

• browser_type

  • Description: Type text into editable element
  • Parameters:
    • element (string): Human-readable element description used to obtain permission to interact with the element
    • ref (string): Exact target element reference from the page snapshot
    • text (string): Text to type into the element
    • submit (boolean): Whether to submit entered text (press Enter after)

• browser_select_option

  • Description: Select option in a dropdown
  • Parameters:
    • element (string): Human-readable element description used to obtain permission to interact with the element
    • ref (string): Exact target element reference from the page snapshot
    • values (array): Array of values to select in the dropdown.

• browser_choose_file

  • Description: Choose one or multiple files to upload
  • Parameters:
    • paths (array): The absolute paths to the files to upload. Can be a single file or multiple files.

• browser_press_key

  • Description: Press a key on the keyboard
  • Parameters:
    • key (string): Name of the key to press or a character to generate, such as ArrowLeft or a

• browser_snapshot

  • Description: Capture accessibility snapshot of the current page (better than screenshot)
  • Parameters: None

• browser_save_as_pdf

  • Description: Save page as PDF
  • Parameters: None

• browser_take_screenshot

  • Description: Capture screenshot of the page
  • Parameters:
    • raw (string): Optionally returns lossless PNG screenshot. JPEG by default.

• browser_wait

  • Description: Wait for a specified time in seconds
  • Parameters:
    • time (number): The time to wait in seconds (capped at 10 seconds)

• browser_close

  • Description: Close the page
  • Parameters: None

Vision Mode

Vision Mode provides tools for visual-based interactions using screenshots. Here are all available tools:

• browser_navigate

  • Description: Navigate to a URL
  • Parameters:
    • url (string): The URL to navigate to

• browser_go_back

  • Description: Go back to the previous page
  • Parameters: None

• browser_go_forward

  • Description: Go forward to the next page
  • Parameters: None

• browser_screenshot

  • Description: Capture screenshot of the current page
  • Parameters: None

• browser_move_mouse

  • Description: Move mouse to specified coordinates
  • Parameters:
    • x (number): X coordinate
    • y (number): Y coordinate

• browser_click

  • Description: Click at specified coordinates
  • Parameters:
    • x (number): X coordinate to click at
    • y (number): Y coordinate to click at

• browser_drag

  • Description: Perform drag and drop operation
  • Parameters:
    • startX (number): Start X coordinate
    • startY (number): Start Y coordinate
    • endX (number): End X coordinate
    • endY (number): End Y coordinate

• browser_type

  • Description: Type text at specified coordinates
  • Parameters:
    • text (string): Text to type
    • submit (boolean): Whether to submit entered text (press Enter after)

• browser_press_key

  • Description: Press a key on the keyboard
  • Parameters:
    • key (string): Name of the key to press or a character to generate, such as ArrowLeft or a

• browser_choose_file

  • Description: Choose one or multiple files to upload
  • Parameters:
    • paths (array): The absolute paths to the files to upload. Can be a single file or multiple files.

• browser_save_as_pdf

  • Description: Save page as PDF
  • Parameters: None

• browser_wait

  • Description: Wait for a specified time in seconds
  • Parameters:
    • time (number): The time to wait in seconds (capped at 10 seconds)

• browser_close

  • Description: Close the page
  • Parameters: None

Computer Use Agent (CUA) Mode

The Playwright MCP also provides a Computer Use Agent (CUA) implementation for high-level, agent-based browser automation. The CUA mode allows for long-running browser sessions controlled by an AI agent.

To use CUA mode, you need to set your OpenAI API key in the .env file (see .env.example).

CUA Tools

• agent_start

  • Description: Start a new agent session with given instructions
  • Parameters:
    • instructions (string, required): The instructions for the agent to follow

• agent_status

  • Description: Check the status of a running agent session
  • Parameters:
    • sessionId (string, required): Session ID returned from agent_start
    • waitSeconds (number, optional): Time in seconds to wait for completion

• agent_log

  • Description: Get the complete log of an agent session
  • Parameters:
    • sessionId (string, required): Session ID returned from agent_start
    • includeImages (boolean, optional): Whether to include images in the log

• agent_end

  • Description: Forcefully end an agent session
  • Parameters:
    • sessionId (string, required): Session ID returned from agent_start

• agent_get_last_image

  • Description: Get the last screenshot from an agent session
  • Parameters:
    • sessionId (string, required): Session ID returned from agent_start

• agent_reply

  • Description: Send a reply to a running agent session to continue the conversation
  • Parameters:
    • sessionId (string, required): Session ID returned from agent_start
    • replyText (string, required): The reply to send to the agent