Playwright MCP

What it is

Playwright MCP is Microsoft’s MCP server for browser automation. Your AI assistant gets a real browser it can open, click, type in and read. Each page reaches the model as structured text from the accessibility tree. No screenshots, no vision model.

1. You ask the assistant to do something on a page
2. It opens the page with browser_navigate
3. browser_snapshot returns the accessibility tree as text
4. It clicks and types with browser_click and browser_type

Who it suits

Anyone whose work runs through web pages and who would rather the assistant did the clicking. The MCP server suits long loops that keep a browser open, such as exploratory testing or tests that repair themselves. For coding agents tight on context, Microsoft’s README points to its separate Playwright CLI with skills, which uses fewer tokens.

If you never work with web pages, skip it.

Where to use it

Each client has its own setup in the README.

• Coding agents: Claude Code, Codex, Gemini CLI, Goose, Cline and opencode
• Editors: Cursor, VS Code with Copilot and Windsurf
• Chat apps: Claude Desktop
• Most other MCP clients: through the standard config below

How to set it up

A single line adds it in most clients.

1. Add the server with your client’s line below, or paste the standard config into its MCP settings.
2. Ask for something small, such as the title of a page you know, to see the browser open.

Claude Code and Codex each add it with one line.

claude mcp add playwright npx @playwright/mcp@latest

codex mcp add playwright npx "@playwright/mcp@latest"

VS Code adds it from its own command line.

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

Most other clients read the standard config.

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

How to use it

Ask for what you want in plain words and the assistant picks the tools. A browser window opens by default, so you can watch each step.

Arguments change how the browser starts:

The README’s config for that clean session:

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": [
        "@playwright/mcp@latest",
        "--isolated",
        "--storage-state={path/to/storage.json}"
      ]
    }
  }
}

Example

You ask Claude Code to check a checkout on your staging site.

Open our staging site, add the blue mug to the cart, go to checkout and tell me whether the total includes tax.

A browser window opens, the mug goes into the cart and checkout loads. The assistant answers from the page as it read it, quoting the line items and the total shown there. Ask for a screenshot and browser_take_screenshot saves one.

For newcomers and experts

If you are new. MCP is a standard way to give an AI assistant extra tools. This one gives it a web browser. You describe the job in plain words and the assistant clicks and types for you, in a window you can watch. Ask it to open a site you know and tell you what is on the page.

For experts. It drives the browser through Playwright and hands the model the accessibility tree instead of pixels, so actions target elements, not screen positions. That is lighter than a vision loop and steadier across layouts. Pages drawn on a canvas are harder, which is what the coordinate tools behind --caps=vision are for. The README lists 72 tools, and 25 of them are on by default.

The default profile persists per workspace, so one client per workspace at a time. To run two side by side, start each with --isolated or give it a separate --user-data-dir. A headless browser closes after an hour without a tool call, a headed one never does, and --idle-timeout changes that.

Good to know

• Not a security boundary: Microsoft says so plainly. The browser reaches anything your computer can.
• Code it runs: two default tools run code. browser_evaluate runs it in the page, and browser_run_code_unsafe in the server process, which the README calls equal to running any code on your machine.
• Files: access stays inside the workspace unless you pass --allow-unrestricted-file-access
• Sign-ins: sites you sign in to stay signed in, in a profile on disk

On Windows that profile sits here:

%USERPROFILE%\AppData\Local\ms-playwright\mcp-{channel}-{workspace-hash}