Browser MCP
Automate your real Chrome browser locally with AI, supporting vision, human-like input, code execution, macros, and watchdogs.
README
Browser MCP
TypeScript MCP server + Chrome extension — automate your real browser with AI (Cursor, Claude, VS Code, Windsurf, …).
Inspired by BrowserMCP/mcp: same architecture (stdio MCP ↔ WebSocket ↔ extension), expanded toolset (vision/SOM, code execution, macros, watchdogs).
┌──────────────────┐ MCP stdio ┌─────────────────────────┐
│ AI client │ ─────────────► │ @browser-mcp/mcp │
│ Cursor / Claude │ ◄───────────── │ (Node.js / TypeScript) │
└──────────────────┘ └───────────┬─────────────┘
│ WebSocket :17373
▼
┌─────────────────────────┐
│ Chrome extension (MV3) │
│ your real profile │
└─────────────────────────┘
No Python. No remote browser farm. Everything runs locally on your machine.
Why this exists
| Cloud browser bots | Browser MCP | |
|---|---|---|
| Profile | Empty / disposable | Your logged-in Chrome |
| Bot detection | Often blocked | Real fingerprint / cookies |
| Stack | Mixed | Pure TypeScript |
| Latency | Network hop | Local WebSocket |
Features
- Vision (Set-of-Mark) — screenshot with numbered interactive elements
- Human-like input — Bezier mouse paths, typed key delays
- Code execution — run JS in the page or Node/Playwright on the host
- Tabs, cookies, storage, network intercept
- Extract data / PDF
- Macros — record & replay flows
- Watchdogs — background condition polling
Repo layout
packages/
shared/ # WS protocol types (shared by MCP + extension)
mcp/ # TypeScript MCP server (stdio + WS hub)
extension/ # Chrome MV3 extension (load unpacked from dist/)
Quick start
1. Install & build
git clone git@github.com:pomoq-dev/browser-mcp.git
cd browser-mcp
npm install
npm run build
2. Load the Chrome extension
- Open
chrome://extensions - Enable Developer mode
- Load unpacked → select
packages/extension/dist - Pin the extension
3. Point your AI client at the MCP server
Cursor (~/.cursor/mcp.json) / Claude Desktop / Windsurf:
{
"mcpServers": {
"browser-mcp": {
"command": "node",
"args": ["/ABSOLUTE/PATH/TO/browser-mcp/packages/mcp/dist/index.js"]
}
}
}
See mcp-config.example.json.
4. Connect the tab
- Open the page you want to automate
- Click the extension icon → Connect (default
127.0.0.1:17373) - Badge turns green ON
The MCP server starts a WebSocket hub on port 17373 and speaks MCP over stdio to the AI client.
Development
npm run build # shared → mcp → extension
npm start # run MCP server
npm run dev # mcp with tsx watch
npm test
npm run typecheck
npm run inspector # MCP inspector UI
npm run build:extension
Extension watch:
npm run watch -w @browser-mcp/extension
Then reload the extension in Chrome.
Tools (overview)
Navigation & tabs: browser_navigate, browser_get_tabs, browser_new_tab, …
Vision: browser_get_visual_state, browser_get_dom_tree, browser_screenshot
Input: browser_click, browser_type_text, browser_drag_and_drop, browser_hover, …
Code: browser_exec_js_page, browser_run_node_playwright, browser_run_node_script
Session: browser_manage_cookies, browser_manage_storage, browser_intercept_network
Data: browser_extract_data, browser_generate_pdf
Macros & watchdogs: browser_record_macro_*, browser_register_watchdog, …
Full list is exposed via MCP tools/list.
Optional: CDP for Playwright tools
browser_run_node_playwright can attach via Chrome DevTools Protocol:
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
--remote-debugging-port=9222 \
--user-data-dir=/tmp/chrome-mcp-profile
Extension mode already uses your normal profile; CDP is only for heavy host scripts.
Security
The agent gets full control of the connected browser profile (clicks, cookies, arbitrary page JS). Run only with trusted local agents.
Credits
Architecture adapted from BrowserMCP/mcp / Playwright MCP ideas: control the user's browser instead of spawning a disposable one.
License
MIT
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。