Browser Use
Browser agents and stealth cloud browsers for AI: run web tasks, fill forms, extract data.
作者提供的说明(英文原文)
MCP server for web page and cross-platform system screenshots (macOS, Linux, Windows)
这三项永远不会合并成一个分数:意见、测量和人气是不同的东西。
我们还没有为它写通俗的介绍。下面是我们从注册中心读取到的事实。
选择你使用的 AI 应用,然后复制设置。
你需要
你用的是哪个应用?
在上方选择你的应用。我们会记住它,下次按钮会显示“安装到 …”。
打开 Claude Desktop → Settings → Developer → Edit Config,把下面的内容粘贴到 claude_desktop_config.json
{
"mcpServers": {
"screenshot": {
"command": "npx",
"args": [
"-y",
"universal-screenshot-mcp"
]
}
}
}在终端中运行这条命令
claude mcp add screenshot -- npx -y universal-screenshot-mcp粘贴到项目中的 .cursor/mcp.json,或 ~/.cursor/mcp.json
{
"mcpServers": {
"screenshot": {
"command": "npx",
"args": [
"-y",
"universal-screenshot-mcp"
]
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
粘贴到项目中的 .vscode/mcp.json
{
"servers": {
"screenshot": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"universal-screenshot-mcp"
]
}
}
}浏览器会先询问是否打开应用,然后应用会显示设置并请你确认。密钥的值会留空,由你自己填写。
这个服务器在你的电脑上运行,而 ChatGPT 只能连接互联网上的服务器。请选择其他应用。
你的选择通过 Cookie 保存在这台设备上,无需账户。
这些设置是根据作者在注册中心的条目生成的。如果 README 的写法不同,请以 README 为准。
它在你的电脑上运行;安装后,工具列表会显示在你的 AI 应用中。
我们检查的是任何人都能看到的信息;我们没有阅读代码。不适用的检查项不计入总数。 · 检查于 2026年11月7日
注册名称已通过 GitHub 账户 sethbang 验证,该账户拥有此仓库
我们还没有从 GitHub 读取这个仓库
我们还没有从 GitHub 读取这个仓库
它在你的电脑上运行,没有互联网端点
在你的电脑上运行,改用本地设置
它在你的电脑上运行;我们只能读取互联网端点的工具列表
与注册中心里这个分类中所有活跃的服务器比较。只有当发布者拥有该仓库时,星标数才计入。
它在 1 项适用的基础检查中通过了 1 项。我们没有阅读代码。在添加密钥或令牌之前,请先看看这个服务器要求哪些权限,并使用只具备所需权限的密钥。
注册中心没有列出必需的密钥,不过有些服务器会在首次使用时要求你登录。
它在你的电脑上运行。你的 AI 应用会在需要时启动它,数据不会经过作者的服务器(除非它自己调用在线服务)。
GitHub 仓库最后更新于–,最新版本(1.5.0)于8天前进入注册中心。
支持 MCP 的应用,例如 Claude Desktop、Claude Code、Cursor 和 VS Code。你的电脑上还需要有 Node.js。
同一分类中有:Browser Use, Chrome DevTools MCP, MCP Server Browser
同一分类中其他作者的服务器,按星标数排序。
如何看卡片:★ GitHub 星标i三大支柱:评价 · 检查 · 使用情况i更新 = 最新版本发布到注册中心的时间i
Browser agents and stealth cloud browsers for AI: run web tasks, fill forms, extract data.
MCP server for Chrome DevTools
MCP server for browser use access
按原文从作者的仓库复制。PopNx MCP 没有审查或背书其内容。HTML、图片和脚本已被移除。链接会离开本站。
An MCP (Model Context Protocol) server that provides AI assistants with screenshot capabilities — both web page capture via Puppeteer and cross-platform system screenshots using native OS tools.
screencapture, Linux maim/scrot/gnome-screenshot/etc., Windows PowerShell+.NET)take_system_screenshot| Platform | Required Tools | Notes |
|---|---|---|
| macOS | screencapture (built-in) | No additional installation needed |
| Linux | One of: maim, scrot, gnome-screenshot, spectacle, grim, or import (ImageMagick) | maim or scrot recommended for full feature support. For window-by-name capture, also install xdotool. |
| Windows | powershell (built-in) | Uses .NET System.Drawing — no additional installation needed |
# Ubuntu/Debian (recommended)
sudo apt install maim xdotool
sudo dnf install maim xdotool
sudo pacman -S maim xdotool
sudo apt install grim
After installing, you can verify your setup with:
npx universal-screenshot-mcp --doctor
This probes the host and prints copy-pasteable install commands for any missing tools, tailored to your detected distro.
npm install -g universal-screenshot-mcp
Or run directly with npx:
npx universal-screenshot-mcp
git clone https://github.com/sethbang/mcp-screenshot-server.git
cd mcp-screenshot-server npm install npm run build
Add the server to your MCP client configuration. For Claude Desktop, edit ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": { "screenshot-server": { "command": "npx", "args": ["-y", "universal-screenshot-mcp"] } } }
Or if installed from source:
{
"mcpServers": { "screenshot-server": { "command": "node", "args": ["/absolute/path/to/mcp-screenshot-server/build/index.js"] } } }
For Claude Code, register the server with the claude mcp add command:
# Project scope (current directory only)
claude mcp add screenshot-server -- npx -y universal-screenshot-mcp
claude mcp add --scope user screenshot-server -- npx -y universal-screenshot-mcp
Or if installed from source:
claude mcp add screenshot-server -- node /absolute/path/to/mcp-screenshot-server/build/index.js
Verify the server registered with claude mcp list, or check live status from inside a session with /mcp.
For Cursor or other MCP clients, consult their documentation for the equivalent configuration.
The server exposes two MCP tools:
take_screenshotCaptures a web page (or a specific element) via a headless Puppeteer browser.
| Parameter | Type | Required | Description |
|---|---|---|---|
url | string | ✅ | URL to capture (http/https only) |
width | number | — | Viewport width (1–3840) |
height | number | — | Viewport height (1–2160) |
fullPage | boolean | — | Capture the full scrollable page |
selector | string | — | CSS selector to capture a specific element |
waitForSelector | string | — | Wait for this selector before capturing |
waitForTimeout | number | — | Delay in milliseconds (0–30000) |
outputPath | string | — | Output file path (default: ~/Documents/screenshots) |
Example prompt:
Take a screenshot of https://example.com at 1920x1080
take_system_screenshotCaptures the desktop, a specific application window, or a screen region using native OS tools. Works on macOS, Linux, and Windows.
| Parameter | Type | Required | Description |
|---|---|---|---|
mode | enum | ✅ | fullscreen, window, or region |
windowId | number | — | Window ID for window mode |
windowName | string | — | App name (e.g. "Safari", "Firefox") for window mode |
region | object | — | { x, y, width, height } for region mode |
display | number | — | Display number for multi-monitor setups |
includeCursor | boolean | — | Include the mouse cursor in the capture |
format | enum | — | png (default) or jpg |
delay | number | — | Capture delay in seconds (0–10) |
outputPath | string | — | Output file path (default: ~/Documents/screenshots) |
| Feature | macOS | Linux | Windows |
|---|---|---|---|
| Fullscreen | ✅ | ✅ | ✅ |
| Region | ✅ | ✅ (maim, scrot, grim, import) | ✅ |
| Window by name | ✅ | ⚠️ X11 + xdotool | ⚠️ best-effort |
| Window by ID | ✅ | ✅ X11 only | ⚠️ HWND |
| Multi-display | ✅ | ⚠️ tool-dependent | ✅ |
| Include cursor | ✅ | ⚠️ tool-dependent | ⚠️ |
| Delay | ✅ | ✅ | ✅ |
Example prompt:
Take a system screenshot of the Safari window
| Variable | Default | Description |
|---|---|---|
SCREENSHOT_OUTPUT_DIR | Documents/screenshots | Default output directory relative to ~ |
ALLOW_LOCAL | false | Set to true to allow screenshotting localhost/127.x.x.x/[::1] (useful for local dev servers) |
Screenshots are saved to ~/Documents/screenshots by default (configurable via SCREENSHOT_OUTPUT_DIR). Custom output paths must resolve to one of these allowed directories:
| Directory | Description |
|---|---|
~/Documents/screenshots | Default output location (configurable) |
~/Desktop/Screenshots | Original default location |
~/Downloads | User downloads folder |
~/Documents | User documents folder |
/tmp | System temp directory |
This server implements multiple layers of security hardening:
| ID | Threat | Mitigation |
|---|---|---|
| SEC-001 | SSRF / DNS rebinding | URLs validated against blocked IP ranges; DNS resolved pre-request with IP pinning via --host-resolver-rules; navigation redirects re-validated |
| SEC-003 | Command injection | All subprocesses use execFile (no shell); app names validated against SAFE_APP_NAME_PATTERN |
| SEC-004 | Path traversal | Output paths validated with fs.realpath() symlink resolution; restricted to allowed directories |
| SEC-005 | Denial of service | Concurrent Puppeteer instances limited to 3 via semaphore; 120 s hard deadline per call, after which the browser is killed and the slot freed |
| SEC-006 | Popup SSRF bypass | Chromium's popup blocker kept on (Puppeteer disables it by default); any popup that still opens is closed |
For full details, see docs/security.md.
| Command | Description |
|---|---|
npm run build | Compile TypeScript to build/ |
npm run watch | Recompile on file changes |
npm test | Unit tests (fast, fully mocked) |
npm run test:integration | Integration tests (real DNS/filesystem) |
npm run test:e2e | E2E tests (real Puppeteer/native tools) |
npm run test:all | All test tiers together |
npm run test:linux | Linux e2e via Docker (requires Docker) |
npm run test:watch | Run tests in watch mode |
npm run test:coverage | Run tests with coverage report |
npm run lint | Lint source with ESLint |
npm run inspector | Launch MCP Inspector for debugging |
src/
├── index.ts # Entry point — stdio transport ├── server.ts # MCP server factory ├── config/ │ ├── index.ts # Static constants (limits, allowed dirs) │ └── runtime.ts # Singleton semaphore, default directory ├── tools/ │ ├── take-screenshot.ts # Web page capture tool │ └── take-system-screenshot.ts # macOS system capture tool ├── types/ │ └── index.ts # Shared TypeScript interfaces ├── utils/ │ ├── helpers.ts # Response builders, file utilities │ ├── screenshot-provider.ts # Cross-platform provider interface + factory │ ├── macos-provider.ts # macOS: screencapture wrapper │ ├── linux-provider.ts # Linux: maim/scrot/gnome-screenshot/etc. │ ├── windows-provider.ts # Windows: PowerShell + .NET System.Drawing │ ├── macos.ts # Window ID lookup via CoreGraphics │ └── semaphore.ts # Async concurrency limiter └── validators/ ├── path.ts # Output path validation (SEC-004) └── url.ts # URL/SSRF validation (SEC-001)
Tests use Vitest in three tiers:
npm test) — Full dependency injection, no real I/O. Fast feedback loop.npm run test:integration) — Real DNS resolution, real filesystem with temp directories, real Puppeteer against a local HTTP server.npm run test:e2e) — Real native screenshot tools. macOS tests run natively; Linux tests run in Docker via npm run test:linux; Windows tests run on the windows-latest CI runner.npm test # Unit tests (~300ms)
npm run test:linux # Linux provider tests in Docker npm run test:all # Everything
npm run inspector
This launches the MCP Inspector connected to your built server, allowing you to invoke tools interactively.
Apache-2.0 — Copyright 2026 Seth Bang