Playwright MCP Guide

What it does

Playwright MCP is Microsoft's MCP server built on top of its Playwright automation engine. It reads a page through its accessibility tree, a structured map of the elements on screen, rather than analyzing a screenshot, and translates that into precise clicks and inputs. Any MCP-capable client can connect to it the same way once Node.js is installed.

This role has the AI open websites, click buttons, and fill out forms itself, taking over work a person would normally do with a mouse and keyboard.

When it fits

It fits automated testing or regression checks where clicks and inputs need to land at the same exact target every run. It also suits teams already running Playwright in a CI pipeline who want to add agent control on the same foundation.

When it doesn't fit

The limitations below come from the vendor's own official documentation, not our evaluation. As of 2026-09-22

Microsoft states plainly that Playwright MCP is not a security boundary. Its allowed-origins and blocked-origins settings are not a safeguard either, and they do not apply to redirects, so trust decisions cannot rest on this setting alone. A persistent profile can only drive one browser at a time, so multiple clients sharing the same workspace will conflict, and the docs also note that coding agents using it will see higher token cost.

• View source at github.com/microsoft

Compare with

ego (lite) has a coding agent take over an entire macOS browser app, while Playwright MCP is a server that connects the same way to any MCP-capable client. The choice comes down to whether you hand browser control to the agent or manage it through code and protocol.

• ego (lite)

Official link

Other tools in this layer

• ego (lite)
• Aside
• Browser Use
• Browserbase + Stagehand