Playwright MCP

Playwright MCP is Microsoft's Model Context Protocol (MCP) server for browser automation, built on the open-source Playwright testing framework. It lets AI agents and assistants open web pages, click, type, navigate and read page content through MCP tool calls.

Its distinguishing design choice is that it works from Playwright's accessibility tree rather than screenshots. The README describes it as fast and lightweight, LLM-friendly because no vision model is required, and deterministic because actions target structured page elements instead of pixel coordinates.

Setup is usually a single entry in an MCP client's configuration that launches the package with npx (@playwright/mcp@latest). It requires Node.js 18 or newer and works with clients including VS Code, Cursor, Windsurf, Claude Desktop, Goose and others. Configuration flags include --headless to run without a visible browser window, --isolated to keep the browser profile in memory instead of saving it to disk, --user-data-dir to choose a profile location, --caps to enable optional capabilities such as vision, PDF and devtools, and --allowed-origins to restrict which origins the browser requests. The documentation states that --allowed-origins is not a security boundary and does not affect redirects.

A persistent browser profile can be used by only one browser instance at a time, so the README advises running additional parallel clients with --isolated or a separate user data directory.

Microsoft also positions Playwright MCP relative to Playwright CLI with skills. It notes that many coding agents favour CLI workflows because they are more token-efficient, avoiding large tool schemas and verbose accessibility trees in the model's context. MCP remains suited to agent loops that benefit from persistent browser state and iterative reasoning over page structure, such as exploratory automation, self-healing tests and long-running workflows.

Playwright MCP is open source and free to use. Typical uses include reproducing UI bugs, checking user flows after a code change and exploratory testing, preferably with an isolated profile and test accounts rather than a personal signed-in browser.