Microsoft's Playwright MCP feeds agents the DOM, not pixels

What it does

This official Microsoft MCP server wraps Playwright so LLMs can drive a browser without ever looking at it. It replaces screenshots with structured accessibility snapshots, letting an agent click, type, and navigate via raw DOM semantics. The target use cases are exploratory automation, self-healing tests, or any long-running workflow where the model needs to iterate against live page state.

The interesting bit

Microsoft is remarkably candid that this MCP flavor is the heavier, more introspective sibling to their CLI+SKILLS tool. The README explicitly warns that verbose accessibility trees burn tokens, making the CLI a better fit for high-throughput coding agents. That kind of “use the other tool instead” honesty is rare in official docs.

Key highlights

• Purely structured data: no vision model required.
• Deterministic tool application avoids the guesswork of pixel-based interaction.
• Supports a wide roster of MCP clients including VS Code, Cursor, Claude Desktop, Goose, and Junie.
• Maintains persistent browser context for iterative, stateful agent loops.
• Requires Node.js 18 or newer.

Caveats

• Token overhead can be significant; the README itself suggests CLI+SKILLS for coding agents with tight context windows.
• The configuration table in the source is truncated, so the full list of supported flags is unclear.

Verdict

Grab it if you are building autonomous agents that need to reason about page structure over time. Pass and use Playwright CLI+SKILLS instead if you are a coding agent optimizing for token efficiency.