Playwright MCP

What it does (plain English)

Playwright MCP gives Claude its own web browser. Claude can open a URL, click buttons, type into fields, pick options from dropdowns, upload files, read console errors and take screenshots. An MCP server is a plug-in that gives Claude a new ability; this one comes from Microsoft’s Playwright team, the people behind the Playwright browser-testing library.

Instead of guessing from screenshots, it reads the page’s accessibility tree: the same structured list of buttons, links and headings that screen readers use. The README describes this as “LLM-friendly” because no vision model is needed and actions land on exact elements rather than pixel positions.

It runs on your computer. By default it opens a visible (“headed”) browser window, so you can watch what Claude does.

Is it for you?

Pick it if you are building a website or web app with Claude and want Claude to check its own work: “go sign up with a test email and tell me what breaks.” It is also handy for reading pages that only fill in after JavaScript runs.

It needs Claude Desktop or Claude Code, because it is a local program. It does not work as a claude.ai web connector.

One honest note from the README itself: if you mainly use a coding agent like Claude Code, Microsoft suggests its newer Playwright CLI with Skills may use fewer tokens, because MCP loads large tool lists and page snapshots into Claude’s context. The MCP version still makes sense for longer, back-and-forth sessions where keeping the browser open matters.

What you can do with it

• Smoke-test a sign-up flow: “Open http://localhost:4321, click Sign up, fill the form with test data, submit, and report any error message or console error.”
• Check a form’s validation: “Try submitting the contact form empty, then with an invalid email, and list what each attempt shows.”
• Compare light and dark mode: the browser_emulate_media tool can switch the colour scheme; ask Claude to screenshot both.
• Debug a slow page: “Load the pricing page and list the network requests that failed.” It has browser_network_requests for this.
• Read a JavaScript-heavy page: “Open this dashboard URL and summarise the table.”

How to set it up

1. Install Node.js 18 or newer (the README’s stated requirement). Check with node --version.

2. Claude Code: run

   claude mcp add playwright npx @playwright/mcp@latest

3. Claude Desktop: open Settings > Developer > Edit Config and add the README’s standard config:

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

   Save the file and restart Claude Desktop.

4. Start a chat and ask: “Open https://example.com and tell me the page title.” A browser window should appear.

5. Optional flags go after the package name in args, for example --headless to hide the window, --isolated to keep each session’s logins separate, or --caps vision,pdf for extra capabilities.

What we tested (2026-10-18): we ran npx -y @playwright/mcp@latest (package version 0.0.83) in a Linux sandbox, completed the MCP handshake and listed 25 tools, including browser_navigate, browser_click, browser_snapshot and browser_fill_form. We then called browser_navigate on https://example.com and got back the page title “Example Domain”. Our container blocks Chromium’s sandbox, so that run needed --headless --isolated --no-sandbox; on a normal Mac or Windows machine you should not need --no-sandbox. Output is saved in our test log.

Watch out for

• Not a security boundary. The README says so plainly. Claude acts in a real browser, so anything you are logged into in that profile, it can reach.
• Logins persist. By default it uses a persistent profile (on macOS under ~/Library/Caches/ms-playwright/), so cookies stay between sessions. Use --isolated if you want a clean slate each time.
• An “unsafe” tool exists. browser_run_code_unsafe runs arbitrary Playwright code; its own description calls it RCE-equivalent (able to run any code on your machine). Approve it only when you know why it is being called.
• Context size. Page snapshots can be long. On big pages, ask Claude to use browser_find to search the snapshot instead of capturing the whole thing.
• Docker is headless-only. The Docker image only supports headless Chromium.

Alternatives

• Firecrawl MCP if you only need page content turned into clean Markdown, not clicks.
• Cloudflare’s Browser Run MCP server (https://browser.mcp.cloudflare.com/mcp) for fetching pages and screenshots from a hosted service rather than your own machine.
• Exa MCP when the job is finding pages by searching, not driving one.