An MCP App that drives a real Chromium browser with Playwright and renders a live, clickable view of the page inside an MCP host such as Archestra. The agent can browse for you, and you can take over in the panel — click, type, scroll, navigate.
Browser tools (browser_navigate, browser_click, browser_type, browser_press,
browser_scroll, browser_back/forward/reload, browser_read_page) drive one
shared Playwright page. The UI panel polls an app-only browser_screenshot tool (~1/s),
maps your clicks back to viewport pixels, and forwards your keystrokes — so the model
and you drive the same page.
npm install
npm run browsers # one-time: download Chromium
npm run build # bundle the UI into dist/mcp-app.htmlnpm run serve # HTTP transport on :3001 (npm run dev = watch mode)
npm run serve:stdio # stdio transportTo use it from an MCP host (e.g. Archestra), build first and point the host's MCP config at the stdio entry:
{
"mcpServers": {
"browser": {
"command": "npx",
"args": ["-y", "tsx", "/absolute/path/to/browser-mcp-app/main.ts", "--stdio"]
}
}
}| Env var | Default | Meaning |
|---|---|---|
HEADLESS |
unset → headed | 1/true runs Chromium headless. |
PORT |
3001 |
HTTP port. |
SCREENSHOT_QUALITY |
60 |
JPEG quality for the live view. |
Built with the create-mcp-app Agent Skill, which covers the MCP Apps SDK patterns
used here (tool + UI-resource registration, app lifecycle, host context, polling). Use
it when extending the server — invoke /create-mcp-app.
- One shared browser page; actions are serialized so calls don't race.
- The live view is a screenshot stream — page pixels never go to the model; only
browser_read_pagereturns text, on request. - It's a real browser: it can log in and submit forms. Don't point it at sites where an unintended click would be costly.