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)
Native Swift macOS automation — 153 tools: AX, EventKit, voice, browser DOM, undo.
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 …”。
链接指向作者自己的文件;我们不托管也不重新打包它。打开文件后,Claude Desktop 会进行安装。
在终端中运行这条命令
claude mcp add mac-control -- npx -y mac-control-mcp粘贴到项目中的 .cursor/mcp.json,或 ~/.cursor/mcp.json
{
"mcpServers": {
"mac-control": {
"command": "npx",
"args": [
"-y",
"mac-control-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": {
"mac-control": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"mac-control-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6日
注册名称已通过 GitHub 账户 AdelElo13 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(0.10.3)于22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 MCPB, 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。
Native Swift MCP server for full macOS automation. 153 tools in one signed .app bundle — no Python, no Node runtime, no Electron. Full list with parameters: docs/TOOLS.md.
Gives any MCP-compatible client (Claude Desktop, Claude Code, Cursor, etc.) the ability to:
Requires macOS 14.0+. Four options, in order of simplicity:
The server is published as an MCP Bundle — a zip with a manifest.json that Claude Desktop reads directly:
.mcpb file. Claude Desktop opens an install dialog.mac-control-mcp and available immediately in new chats.It's also listed on the official MCP Registry as io.github.AdelElo13/mac-control-mcp, so any MCP client that supports the registry will find it by searching for "mac-control".
For clients that are configured with a command line rather than a file path. The npm package mac-control-mcp is a thin launcher: it has no runtime dependencies, and on install it downloads the notarized MacControlMCP.app for that exact version from this repo's GitHub release, verifies it against the published .sha256, refuses any archive entry that escapes the install directory, and re-checks the bundle with codesign --verify --deep --strict, spctl --assess --type execute and a pinned Team ID (A3W973JZ49). The first run therefore pulls ~3 MB; later runs are local.
Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": { "mac-control-mcp": { "command": "npx", "args": ["-y", "mac-control-mcp"] } } }
Claude Code:
claude mcp add mac-control-mcp -- npx -y mac-control-mcp
Cursor — ~/.cursor/mcp.json:
{
"mcpServers": { "mac-control-mcp": { "command": "npx", "args": ["-y", "mac-control-mcp"] } } }
ChatGPT (developer mode, local MCP connector) — mcp.json:
{
"mcpServers": { "mac-control-mcp": { "command": "npx", "args": ["-y", "mac-control-mcp"], "transport": "stdio" } } }
Two flags are handled by the launcher itself instead of being forwarded to the binary:
npx -y mac-control-mcp --version # version of the bundled .app
npx -y mac-control-mcp --app-path # path to MacControlMCP.app
--app-path is what you need when granting macOS permissions by hand. Note that TCC grants go to the MCP host application that launches this process (Claude Desktop, Cursor, your terminal), not to npx or to the script — run the permissions_status tool once from your client to see which grants are missing and which app they belong to.
MAC_CONTROL_MCP_SKIP_DOWNLOAD=1 skips the download (CI, sandboxed builds); HTTPS_PROXY and NO_PROXY are honoured.
If you don't use Claude Desktop or want manual control:
MacControlMCP.app to ~/Applications/.{
"mcpServers": { "mac-control-mcp": { "type": "stdio", "command": "/Users/you/Applications/MacControlMCP.app/Contents/MacOS/MacControlMCP" } } }
Add that block to ~/Library/Application Support/Claude/claude_desktop_config.json (Claude Desktop) or ~/.claude.json → mcpServers (Claude Code).
Verify the download with the published SHA-256:
shasum -a 256 MacControlMCP-v<version>-macos-universal.tar.gz
For contributors or if you want to tweak the code. Requires Swift 6 / Xcode 16+:
git clone https://github.com/AdelElo13/mac-control-mcp.git
cd mac-control-mcp ./scripts/build-bundle.sh
Produces ~/Applications/MacControlMCP.app/Contents/MacOS/MacControlMCP. Without a Developer ID cert in your keychain it'll fall back to ad-hoc signing (works for local use, TCC grants reset on every rebuild).
To re-sign + re-notarise an existing Apple Developer account:
# one-time: store notary credentials in keychain
xcrun notarytool store-credentials "mac-control-mcp" \ --apple-id "you@example.com" --team-id "XXXXXXXXXX"
NOTARIZE_PROFILE=mac-control-mcp ./scripts/build-bundle.sh
Selected categories — this table is a curated subset, not the full list. For the complete, auto-generated tool reference (every tool, its description, and its required/optional parameters, regenerated straight from the tool registry so it can't drift), see [docs/TOOLS.md](docs/TOOLS.md).
| Category | Tools |
|---|---|
| Permissions | permissions_status, request_permissions |
| Accessibility | find_element(s), query_elements, list_elements, get_ui_tree, ax_tree_augmented, get_element_attributes, set_element_attribute, read_value, perform_element_action, wait_for_element, scroll_to_element |
| App lifecycle | list_apps, launch_app, activate_app, quit_app, force_quit_app, wait_for_app, focused_app |
| Windows | list_windows, focus_window, move_window, resize_window, set_window_state, wait_for_window, move_window_to_display |
| Input | click, mouse_event, drag_and_drop, scroll, type_text, press_key, press_key_sequence, key_down, key_up, convert_coordinates |
| Menus | click_menu_path, list_menu_paths, list_menu_titles |
| Browser | browser_list_tabs, browser_get_active_tab, browser_navigate, browser_new_tab, browser_close_tab, browser_eval_js, browser_dom_tree, browser_visible_text, browser_iframes |
| Screen | capture_screen, capture_screen_v2, capture_window, capture_display, ocr_screen |
| Clipboard | clipboard_read, clipboard_write, clipboard_clear |
| Spotlight | spotlight_search, spotlight_open_result |
| System | set_volume, set_dark_mode, list_displays |
| File dialogs | file_dialog_set_path, file_dialog_select_item, file_dialog_confirm, file_dialog_cancel, wait_for_file_dialog |
| Apple apps | mail_send, imessage_send, imessage_list_recent, calendar_create_event, calendar_list_events, reminders_create, reminders_list, contacts_search |
| Voice & recording | speech_to_text, text_to_speech, audio_record, record_screen |
| Undo | undo_last_action, undo_peek |
153 tools total — see docs/TOOLS.md for the complete, generated list.
Several tools overlap in purpose. Based on their actual implementations:
Screen capture — `capture_screen` vs `capture_screen_v2` vs `capture_window` vs `capture_display`
capture_screen — main display (or a rectangular region within it) to a PNG file, returns the file path + width/height. Simple, synchronous, no size limits applied.capture_screen_v2 — main display only, but stores the PNG as a content-addressed artifact under ~/.mac-control-mcp/artifacts/ .png and returns {content_ref, bytes, sha256}. Defaults to inline=false and downscales to max_dimension/max_bytes (4000px / 4MB by default) specifically to avoid blowing up an MCP client's context window with a huge inline image (referenced in its own description as a fix for claude-code issues #13383/#45785). Prefer this one when the image is going back through an LLM context rather than straight to disk.capture_window — screenshots one specific window by PID (optionally filtered by title_contains), not the whole display.capture_display — screenshots one specific physical display by its index from list_displays, for multi-monitor setups.Accessibility tree / element lookup — `find_element` vs `find_elements` vs `query_elements` vs `list_elements` vs `get_ui_tree` vs `ax_tree_augmented`
list_elements — lists actionable elements for a PID (a flat, practical "what can I click" view).find_element — returns the *first* element matching a role/title filter for a PID.find_elements — same matching as find_element, but returns *all* matches, not just the first.query_elements — regex search over role/title/value (falls back to case-insensitive substring on invalid regex); use when a plain role/title filter isn't precise enough.get_ui_tree — walks the *full* accessibility tree of a process, including containers, and assigns stable element IDs for follow-up calls. Heavier than list_elements/find_element(s), but the only one that gives you structure/hierarchy.ax_tree_augmented — like get_ui_tree, but does one OCR pass and geometrically joins OCR-derived labels onto AX nodes that have no native label. Use it specifically for Electron/Chromium/Canvas apps where the native AX tree is sparse and unlabeled; it's the expensive option (an OCR pass), so reach for get_ui_tree first and fall back to this when nodes come back unlabeled.capture_*, ocr_screen) validate output_path via a strict allow-list — only the user-scoped temp dir (NSTemporaryDirectory()) and ~/Desktop, ~/Documents, ~/Downloads, ~/Pictures are accepted. Symlinks at the target path are rejected to prevent redirection. /tmp is deliberately excluded because it's shared across users and opens a TOCTOU window.browser_eval_js wraps user code in (0, eval)(…) via JSON.stringify, so quotes/newlines/unicode can't break out of the wrapper.AXSecureTextField never returns its value or length, and every write route is refused with reason: secure_field: element writes, text_* edits, act, untargeted type_text, and text keys or Cmd+V via press_key, press_key_sequence and key_down while such a只显示了开头部分;其余内容请在仓库中阅读