Playwright MCP 🔌 MCP server Free

What it does

Playwright MCP is Microsoft's MCP server for browser automation, built on the Playwright testing library. Instead of screenshots, it gives the agent structured accessibility snapshots of the page, so it can find and use elements by role and name without a vision model. For vibe coders, it is a common way to let an agent open the app it just built, click through a flow and check that the flow works. The project is Apache-2.0 licensed. Microsoft's own README says that a coding agent short on context may do better with its sibling, Playwright CLI plus Skills. MCP makes more sense for longer sessions where the browser needs to stay open.

Tools it gives your agent

• Navigation, clicking, typing, hovering, drag and drop, form filling, dropdown selection, file uploads and dialog handling
• browser_snapshot for the accessibility tree, plus screenshots, console messages and network requests
• Tab management and waiting for text or elements
• Optional tools you switch on with --caps: coordinate-based mouse control (vision), PDF export, DevTools-style tracing and video recording, and cookie or storage control

Works with

It runs locally on Node.js 18 or newer and works with any MCP client. The README includes setup steps for Claude Code, Claude Desktop, Codex, Copilot CLI, Cursor, VS Code, Gemini CLI, Cline, Goose, Kiro, Junie, Amp and Factory. You can also run it as a standalone HTTP server with --port, or as a headless Chromium container using the mcr.microsoft.com/playwright/mcp image.

How to connect

claude mcp add playwright -- npx @playwright/mcp@latest

Standard config for other clients:

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

Cost and limits

It is free and open source, and you need no account or API key. The main cost is context: snapshots and the long tool list take up tokens on every run.

Safety notes

The README states that Playwright MCP is not a security boundary. By default the browser keeps a persistent profile, so logins carry over between runs. Use --isolated to keep the profile in memory instead. The --allowed-origins and --blocked-origins options help keep the agent on target, but they are not security controls. The browser_run_code_unsafe tool runs arbitrary JavaScript in the server process, so only allow it when you trust the prompt.