Source: Playwright MCP README

Microsoft's Playwright MCP server: browser automation for LLMs via accessibility-tree snapshots rather than screenshots, with opt-in vision mode and security caveats.

Contents
  • What: open-source project README
  • Author/publisher: Microsoft
  • Date: living document (2026); read 2026-11-14
  • Original: microsoft/playwright-mcp

Summary

An MCP server exposing Playwright browser control as tools. The model acts on structured accessibility snapshots, so no vision model is required and actions target elements deterministically.

Key claims

  • Accessibility tree instead of pixels avoids the ambiguity of screenshot-based approaches.
  • Core tools: navigate, snapshot, click, type, fill form, select option, evaluate, wait, tabs, console and network inspection; opt-in capability groups (storage, network mocking, PDF, devtools, testing).
  • --caps=vision adds coordinate-based mouse tools for elements the snapshot doesn’t expose.
  • For coding agents, Playwright CLI with skills may use fewer tokens; MCP suits persistent browser state and iterative exploration.
  • Not a security boundary: browser_run_code_unsafe is RCE-equivalent, and origin allow/block lists don’t cover redirects.

Feeds

Computer-use agents

Suggest changes →