What Playwright's MCP actually does
Microsoft's Playwright MCP lets an agent drive a local browser from an accessibility tree, with Cursor and Claude install steps and nearby PlugMCP listings.
Playwright MCP is Microsoft's browser server for agents. It runs a real browser on your machine and talks to the page through Playwright's accessibility tree, not through screenshots. The agent gets a structured snapshot, picks a target, and clicks or types. No vision model is required for the basic loop.
This week the server is still moving. Appsgit's MCP trending list for 31 Oct to 7 Nov 2026 includes Playwright MCP among the weekly star gainers, and MCP Trends ranks it near the top of its seven-day momentum list. On PlugMCP it is the Microsoft listing under browser automation: https://plugmcp.ai/mcp/microsoft-playwright-mcp.
What it is good for
Use it when the agent needs a live tab: fill a form, click through a flow, read what is on the page after JavaScript ran, or drive a headed browser you can watch. The README's core tools cover navigation, snapshots, clicks, typing, forms, tabs, console and network, screenshots, and waiting. Optional flags add PDF export, coordinate clicks, and DevTools-style recording.
The package on npm is @playwright/mcp (checked at 0.0.84 when this was written). The package and the GitHub README ask for Node.js 18 or newer. Microsoft also ships a Playwright CLI with skills. The README says coding agents that care about token cost may prefer that CLI. MCP still fits when you want a persistent browser session and iterative work on the same page.
Install in Cursor
Node.js 18 or newer. In Cursor, open Settings, then MCP, then add a server. Use a command of npx with args @playwright/mcp@latest, or paste this into your MCP config:
{
"mcpServers": {
"playwright": {
"command": "npx",
"args": ["@playwright/mcp@latest"]
}
}
}
There is also a one-click install button on the Microsoft README. After it connects, ask the agent to open a page and list what it can see. You should get a snapshot, not a blank tool list.
Install in Claude
Claude Code:
claude mcp add playwright npx @playwright/mcp@latest
Claude Desktop uses the same standard config as Cursor: command npx, args @playwright/mcp@latest in claude_desktop_config.json, then fully quit and restart the app.
If you need a headed browser on a machine without a display, the README says to start the server with a port and point the client at the HTTP URL instead of stdio:
npx @playwright/mcp@latest --port 8931
{
"mcpServers": {
"playwright": {
"url": "http://localhost:8931/mcp"
}
}
}
Docker is available as mcr.microsoft.com/playwright/mcp. The image notes say it supports headless Chromium only.
Limits to know
- Playwright MCP is local. Your agent starts a process. It is not a remote URL on the PlugMCP listing.
- It is not a security boundary. The README says that plainly. The browser can reach whatever the machine can reach.
- A persistent profile is the default. Concurrent clients on the same workspace can collide. Use --isolated or a separate user data directory if you run more than one client.
- Accessibility snapshots are strong for ordinary pages and weaker when the UI is canvas-only or hostile to assistive tech. Prefer snapshot tools for clicks and fills. Screenshots exist, but the README tells you not to drive actions from them.
- Microsoft notes that for high-throughput coding agents that also hold a large codebase in context, the Playwright CLI plus skills can be cheaper on tokens than loading MCP tool schemas and full trees.
Alternatives on PlugMCP
Same shelf, different jobs:
- Chrome DevTools MCP when you care about performance traces, network detail, and Chrome-specific debugging:
https://plugmcp.ai/mcp/chromedevtools-chrome-devtools-mcp - Scrapling when the job is scrape-and-extract from a local Python process more than a long interactive browser loop:
https://plugmcp.ai/mcp/d4vinci-scrapling - Firecrawl when you want a remote scrape and search URL rather than a local browser:
https://plugmcp.ai/mcp/firecrawl-firecrawl-mcp-server
On the PlugMCP page
The listing is https://plugmcp.ai/mcp/microsoft-playwright-mcp. Transport is stdio. Status is local. The install block is the publisher's local command, not an mcp.plugmcp.ai proxy. Hosting a copy on PlugMCP is a separate action if you want keys and logs on a server you operate.
Search more browser servers from the directory MCP at https://plugmcp.ai/mcp-server, or open the browser automation category at https://plugmcp.ai/categories/browser-automation.
