Playwright MCP — drive a local browser from an agent

Playwright MCP is Microsoft's MCP server for browser automation (@playwright/mcp, Apache-2.0). It lets an agent open pages and click or type on them, working from an accessibility snapshot rather than screenshots: every action targets an element ref from the latest snapshot the server hands back.

Use it when a task needs a real page — open a URL, read what is on it, fill a form, or check that a control is there. The MCP client must be able to start a local stdio server, and Node.js 18+ is required. Setup is confirming the package, installing the browser it launches, and pointing the client at the server; --headless matters on a machine with no display, since headed is the default.

The core loop is three tools: browser_navigate with a URL (the result includes a snapshot), browser_find with text to locate nodes on a large page, and browser_click / browser_type with a target ref from that snapshot. browser_snapshot re-captures the tree, with a depth option when the full tree is too large.

The gotchas are operational: a ref is only valid from the latest snapshot, so any click or navigation means re-snapshotting before the next target; screenshots are for looking, not for aiming. Extra capability groups (vision, pdf, devtools, network, storage, testing, config) stay off until enabled with --caps. The server binds to localhost and is not a security boundary, file access stays inside the workspace roots, and browser_run_code_unsafe — arbitrary JavaScript in the server process — is to be left alone.