Microsoft Playwright MCP

About

Microsoft's official Playwright MCP server (35,900+ GitHub stars, npm `@playwright/mcp`) drives a real browser through Playwright's accessibility tree rather than screenshots, so it needs no vision model and its responses are structured text an LLM can act on deterministically. Requires Node.js 18+; `claude mcp add playwright npx @playwright/mcp@latest` is the whole install for Claude Code, and the same stdio block works in Cursor, Claude Desktop, Windsurf, Codex, Goose, VS Code, Copilot, Kiro and LM Studio. Roughly 70 tools ship across eight groups: core automation (browser_snapshot, browser_find, browser_click, browser_fill_form, browser_evaluate, browser_console_messages, browser_network_requests), tab management, and six opt-in capability groups behind --caps — network request mocking and offline simulation, cookie/localStorage/sessionStorage management, devtools tracing and video, coordinate-based mouse input via vision, PDF export, and testing assertions with browser_generate_locator and the verify_* tools. By default it runs headed with a persistent profile, so logins survive between sessions; that profile is keyed to the client workspace and can only be used by one browser instance at a time, so a second concurrent client needs --isolated or its own --user-data-dir. File-system access is restricted to workspace roots and file:// navigation is blocked unless --allow-unrestricted-file-access is passed, while --allowed-origins and --blocked-origins are explicitly documented as not a security boundary. Every CLI flag has a matching PLAYWRIGHT_MCP_* environment variable. Microsoft now points coding agents at the Playwright CLI with skills instead, on token-efficiency grounds, and recommends MCP for agentic loops that need persistent browser state.

Installation

npx @playwright/mcp@latest

Installing Playwright MCP

1. 1.Check the runtime

   Node.js 18 or newer. Everything below runs through npx, so there is nothing to install globally.

2. 2.Claude Code

   One command, no arguments needed for a default headed persistent-profile setup.

   shell

   claude mcp add playwright npx @playwright/mcp@latest

3. 3.Anything with a JSON config

   Cursor, Claude Desktop, Windsurf, Goose, Codex, LM Studio, Kiro, Copilot and others take the same stdio block. Flags go in args after the package name.

   json

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

4. 4.Decide persistent or isolated before you run two clients

   The default is a persistent profile, so logins survive across sessions. It lives in ~/Library/Caches/ms-playwright/mcp-{channel}-{workspace-hash} on macOS, ~/.cache/ms-playwright/... on Linux and %USERPROFILE%\AppData\Local\ms-playwright\... on Windows, where the workspace hash comes from the MCP client's workspace root — so different projects get separate profiles automatically. A persistent profile can only be used by one browser instance at a time.

5. 5.Run several clients in parallel

   Start each additional client with --isolated, or point it at a distinct --user-data-dir. Isolated sessions keep the profile in memory and lose all storage state when the browser closes; seed them with --storage-state pointing at a Playwright storage-state file if they need to start logged in.

   json

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

6. 6.Turn on the capability groups you need

   Core automation, tab management and browser installation are always on. Everything else is opt-in behind --caps: config, network, storage, devtools, vision and pdf, plus testing for assertion tools. Note that the README's options table lists only vision, pdf and devtools while its tool sections document the rest — trust the tool sections, and verify with a tools/list against your own install.

   shell

   npx @playwright/mcp@latest --caps=network,storage,testing

The tool surface, by group

Roughly 70 tools across eight groups. The ones worth knowing by name are the ones that change how you prompt.

What people use it for

Which one should you use?

Two different projects answer to "the Playwright MCP server". They are not the same code.

Categories

Works With

Frequently Asked Questions