Playwright

Microsoft’s Playwright MCP server drives a real browser for an agent, but deliberately avoids the usual screenshot-and-vision-model loop. Instead it operates on structured accessibility snapshots of the page — the same tree a screen reader would use — which makes its actions more token-efficient and considerably more deterministic than pixel-based approaches.

That combination makes it well suited to exploratory automation (an agent figuring out how to complete a flow it hasn’t seen before), self-healing test suites that need to tolerate minor UI drift, and long-running autonomous workflows where reliability matters more than visual fidelity. Configuration is almost entirely optional — PLAYWRIGHT_MCP_BROWSER and PLAYWRIGHT_MCP_HEADLESS cover the two settings most people touch, and everything else has a sane default.

Client support is unusually broad: beyond the major agent clients, it’s documented working with editors like VS Code and Cursor, CLI tools, and a long tail of other MCP hosts — a reasonable default choice whenever a task calls for a real browser rather than a static-page fetch.