mcp-cui-chrome
Controls a locally installed Google Chrome window through tools for navigating pages, taking page snapshots, clicking, typing, pressing keys, hovering, dragging, scrolling, managing tabs, going back, and capturing screenshots. It uses a separate persistent Chrome profile and is not tied to any specific website.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@mcp-cui-chrome 打开 Chrome,去百度搜索今天的天气"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
mcp-cui-chrome
通用的本地 Chrome 操作 MCP。它不针对任何一个网站,也不内置某个业务的填写流程。服务说明会告诉模型:用户说打开 Chrome、打开网页、登录、点击或看页面时,用这套工具,不要改用 IDE 里的内置浏览器。窗口左下角有黄色 cui-chrome 标签。
浏览器是这台电脑上已经安装的 Google Chrome,配置目录单独放在 %LOCALAPPDATA%\mcp-cui-chrome\chrome-profile,不会占用日常使用的那个窗口。在这个窗口里登录过的网站,下次还会保持登录。
多个 Cursor 会话共用这一个窗口。先连上的会话负责打开它,后连上的会话接到同一个窗口,不再另开一个。两边的点击、输入会排队。后连上的会话如果调用关闭,只断开自己,窗口留着。
每一步都有说明
每个动作都会先停一下,再动手:
开始前,窗口左下角的黄底标签写成
cui-chrome 准备:…,工具返回值的第一行也是这句。做完后,标签和返回值都改成
结果:成功。…或结果:失败。…。失败时带上原因。
页面上每个元素都会标出控件类型和对应操作,例如「操作:输入文字」「操作:点开后点选一项,不要输入文字」「操作:点击」「操作:拖动」。输入工具只接受文本框。下拉框要先点开,再点里面的选项。
Related MCP server: chrome-agent-mcp
环境
Node.js 20 或更高
已安装的 Google Chrome
安装
本机已安装 Google Chrome 时,在 MCP 配置里这样接:
{
"mcpServers": {
"cui-chrome": {
"command": "npx",
"args": ["-y", "mcp-cui-chrome"]
}
}
}改完后重载 MCP。日志走 stderr,标准输出只留给协议。
从源码开发:
npm install
npm run build工具
模型先看页面,再按编号操作。编号形如 e1、e3,页面一变就要重新读取。
工具 | 作用 |
| 打开网址。相同地址的标签已存在就切过去,否则新建标签 |
| 标题、网址、可操作元素编号、正文。有遮挡层会先写出来 |
| 刷新当前页,相当于 F5。本该有的内容没有时用一次 |
| 沿鼠标轨迹移动后点击 |
| 逐字输入,可选先清空、输入后回车 |
| 按键,例如 Enter、Escape |
| 悬停,不点击 |
| 按住一个元素拖到另一个元素,或按像素偏移拖动 |
| 分段滚轮 |
| 列出、新建、切换、关闭标签 |
| 后退 |
| 当前视口截图 |
| 关掉自己打开的 Chrome;接到别人的窗口时只断开连接 |
点击、输入和滚动都带随机停顿。鼠标走贝塞尔曲线,落点在元素内部的随机位置,不是每次都点正中心。
环境变量
变量 | 作用 |
| 换一个配置目录 |
| Chrome 不在默认位置时,填 |
| 设为 |
| 给后启动的会话留的本机调试端口。不设或设为 |
能力边界
启动参数去掉了自动化开关,并用补丁版 Playwright 遮住常见的调试协议痕迹。这能降低「一眼看出是脚本」的概率。依赖 IP 信誉或验证码的网站仍然可能拦住访问。这个服务不处理验证码。
本地冒烟
npm run smoke会弹出 Chrome,打开 example.com,点其中的链接,再打开检测页看 navigator.webdriver。
Available Tools
13 toolsbrowser_click拟人点击A
通用:按快照编号点击。开始前显示要点的元素,结束后显示是否成功、页面有没有跳转。落点被别的内容挡住时会失败,不点穿过去。
Inputs: ref (required), button.
browser_close关闭浏览器A
通用:关掉这次会话自己打开的 Chrome。如果窗口是别的会话打开的,只断开连接,窗口留着。
Inputs: no named parameters.
browser_drag拖拽A
通用:按住一个元素,拖到另一个元素上,或按像素偏移拖动。用于滑块、排序和需要拖动的控件。
Inputs: from (required), dx, dy, to.
browser_go_back后退A
通用:当前标签后退一页。开始前说明,结束后显示是否退回到上一页。
Inputs: no named parameters.
browser_hover悬停A
通用:把鼠标移到编号对应的元素上,不点击。用来展开需要悬停才出现的菜单。
Inputs: ref (required).
browser_press按键B
通用:在当前焦点按一个键。开始前显示键名,结束后显示是否成功。
Inputs: key (required).
browser_reload刷新页面A
通用:刷新当前标签,相当于 F5。快照里本该出现的按钮或正文没有时用一次,然后再读页面。不要连续反复刷新。
Inputs: no named parameters.
browser_screenshot截图A
通用:截取当前视口。开始前在页面上说明,结束后返回图片并标明成功或失败。
Inputs: no named parameters.
browser_scroll滚动B
通用:在当前页面分段滚动。开始前显示方向,结束后显示是否成功。
Inputs: direction (required), pixels.
browser_snapshot读取页面A
用户要看当前 Chrome 页面时用这个。读取标题、网址、可操作元素编号和正文。若本该有的内容不在,用 browser_reload 刷新一次再读。快照若写了遮挡,先关掉那一层或按 Escape。
Inputs: no named parameters.
browser_tabs标签页C
通用:列出、新建、切换或关闭标签。开始前说明要做什么,结束后说明是否成功。
Inputs: action (required), url, index.
browser_type拟人输入A
通用:只对文本框逐字输入。下拉框、选项、按钮会直接失败,并说明该用点选还是点击。
Inputs: ref (required), text (required), clear, submit.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 13 tools
Each tool maps to a clearly distinct browser action (navigate, click, type, press, hover, drag, scroll, screenshot), and descriptions explicitly cross-reference siblings (e.g. snapshot vs reload, tabs vs navigate). Minor overlap remains between browser_navigate and browser_tabs for opening tabs, and between browser_snapshot and browser_screenshot, but the descriptions differentiate them well enough.
All 13 tools use the same browser_ prefix with snake_case, and the suffix is either a verb (navigate, click, type, press, hover, drag, scroll, reload) or a clearly paired noun (tabs, snapshot, screenshot, close). The pattern is predictable and readable throughout.
13 tools is well within the ideal 3–15 range and every tool corresponds to a real, non-redundant browser control primitive. Nothing feels padded or missing at the count level.
The surface covers the full browsing lifecycle: navigation, history, tabs, DOM reading, all common input/pointer interactions, screenshots, and session cleanup. Gaps exist for file upload, JS evaluation/console or network inspection, and explicit wait-for-condition, but core agent workflows are workable.
Related MCP Connectors
Give AI agents a real cloud Chrome: open sessions, navigate, read pages, click, type and screenshot.
Let AI agents use your real, signed-in Chrome: tabs, clicks, forms, screenshots and video.
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
Let your AI agent operate any website through a real, logged-in browser session.
Related MCP Servers
- AlicenseBqualityDmaintenanceEnables AI agents to directly control your real Chrome browser with full context including login sessions, cookies, and open tabs. It provides tools for page scanning, JavaScript execution, CDP control, screenshots, and physical mouse/keyboard input for authentic browser automation.20244MIT
- AlicenseBqualityDmaintenanceEnables AI agents to fully control Google Chrome: navigate, click, fill forms, inspect DevTools, and manage tabs with parallel execution and session isolation.249 npmMIT
- AlicenseNot gradedqualityDmaintenanceEnables AI assistants to control a Chrome browser through natural language, supporting navigation, interaction, tab management, and data extraction.MIT
- AlicenseNot gradedqualityAmaintenanceEnables AI assistants to drive the user's real, already-logged-in Chrome browser locally, so they can read and navigate pages, click, type, fill forms, take screenshots, and inspect cookies, tabs, console, and network traffic without headless browsers or remote-debugging ports.1 npmISC