playwright-mcp

Playwright is Microsoft's browser automation library, the one half the web's end-to-end tests run on. This MCP server hands it to the agent: the model can open a page, click, fill a form, wait for a load and read the result.

The key difference from "download the HTML and parse it": the agent works against the page's accessibility tree — the same structure a screen reader consumes — not the source. So it sees a button as a button and a field as a field, and it does not have to guess selectors from class names that change on the next deploy.

What that buys you. Checking that the form on your landing page actually submits. Walking the sign-up flow after a change to confirm nothing broke. Pulling data from a site that renders nothing without JavaScript. The agent does all of it itself, showing its steps.

One command installs it into Claude Code, it runs over stdio, and it also connects to Claude Desktop and Cursor. A security note worth stating plainly: a browser under an agent's control walks the real internet as you — do not hand it sessions with access to money or mail until you can see exactly which steps it takes.