Reference claude in chrome vs playwright mcp
name: claude-in-chrome disabled globally, Playwright MCP is now default web tool description: Global config change: claudeInChromeDefaultEnabled=false, Playwright MCP registered at user scope as the default headless browser tool type: reference
On 2026-07-05, claudeInChromeDefaultEnabled in ~/.claude.json was set to false (backup at ~/.claude.json.bak.chrome-playwright), and the playwright MCP server (npx -y @playwright/mcp@latest) was registered at user scope across all projects, with Chromium pre-installed.
Why: claude-in-chrome was auto-injecting ~30 browser tools (navigate, computer, read_page, etc.) plus a "use these for web tasks" instruction block into every session, causing agents to default to it even though the Chrome extension bridge was disconnected (cachedChromeExtensionInstalled = false), producing repeated "browser extension isn't connected" failures. Justin wanted a headless, self-contained default instead, with chrome kept available on-demand rather than always-on.
How to apply:
- Playwright MCP is now the default web-automation tool for new sessions (headless, no extension/login dependency) — reach for it first for scraping, screenshots, DOM reads, testing.
- claude-in-chrome no longer auto-loads; to use it in a session, flip claudeInChromeDefaultEnabled back on via /config or set the flag true for that session. Use it only when the task needs Justin's real, logged-in Chrome session (an authed site headless can't reach).
- Vercel MCP was considered and explicitly rejected: Vercel is deployment-management only and Justin's sites deploy to Cloudflare Pages, not Vercel — do not suggest Vercel MCP for browser/web automation.
- /screenshot and /preview-site skills already used Playwright as a local CLI independent of this MCP change and are unaffected.
- Takes effect on next Claude Code restart, not the session where it was changed.
[auto-memory session 2e05b6d6-3ee2-4c56-bdcd-14a1528c0c98, confidence 0.85, mode direct]