auto-browser

通过 Chrome Debug 模式（CDP）自动化操作真实浏览器——导航、点击、填表、提取数据、截图、执行多步流程。当用户说"帮我在网页上操作"、"打开浏览器"、"帮我点"、"帮我填"、"帮我抓取"、"帮我截图"、"auto-browser"时使用。

Security Scan

The skill's instructions broadly match a browser-automation purpose, but it contains several unexpected or risky operational steps (auto-starting Chrome, editing a user config to add an npx-based tool runner, and implicit reliance on local tooling) that are not declared in the metadata and merit user attention.

Purpose & Capability

The skill's name/description (real-browser automation via CDP) matches the runtime instructions: navigation, clicks, form-filling, snapshots, JS evaluation, screenshots. However the metadata claims no required binaries or env vars while SKILL.md expects curl, nohup (shell), a local Chrome installation, and npx/@playwright/mcp — a mismatch between declared requirements and actual operational needs.

Instruction Scope

Instructions tell the agent to check http://127.0.0.1:9222 and, if no response, automatically start Chrome using a hardcoded macOS path and nohup (without asking the user). It also instructs modifying ~/.cursor/mcp.json to add a command that will run npx @playwright/mcp. The skill allows executing arbitrary page JS (browser_evaluate) and reading/extracting page data and writing results to disk. These are within a browser-automation use case but the automatic process-start and config-edit behaviors expand scope and pose privacy/permission risks.

Install Mechanism

There is no formal install spec, but the skill instructs to configure MCP to call npx @playwright/mcp (which will fetch code from npm at runtime). That means the agent may cause network installs and execute third-party code via npx. The skill also assumes standard shell tools (curl, nohup) and direct execution of the Chrome binary; these implicit dependencies are not declared, increasing risk.

Credentials

The skill declares no required credentials, yet connecting to a local CDP endpoint (127.0.0.1:9222) gives powerful access to the user's browser session (cookies, localStorage, active logins). The SKILL.md does partially mitigate by requiring user confirmation for sensitive write/irreversible actions (password fills, submits, payments), but the ability to automatically start and control a browser without explicit per-action consent is high privilege and should be treated cautiously.

Persistence & Privilege

The skill directs adding a 'playwright-cdp' entry to ~/.cursor/mcp.json (a persistent change to MCP configuration). Writing its own agent config is expected for this type of tool, but the instruction recommends doing so programmatically and then asking the user to reload — users should be aware this modifies a local config file.

What to consider before installing

This skill can control a real Chrome instance and modify your Cursor MCP config and may trigger network installs via npx. Before installing or using it, consider: 1) The SKILL.md will try to start Chrome automatically (it uses a hardcoded macOS binary path) — decide if you want the agent to run system commands without asking. 2) CDP access (localhost:9222) can expose cookies, tokens, and any page data; only use this skill when you trust the agent and have no sensitive pages open. 3) The skill instructs adding an npx/@playwright/mcp command to ~/.cursor/mcp.json, which will cause transient downloads from npm when invoked — review that package and prefer manual installation if you want to vet the code. 4) The metadata does not declare required tools (curl, nohup, npx, Chrome), so expect to provide or permit these at runtime. If you proceed, require explicit confirmations for sensitive actions (password fills, submits, payments) and review any config edits before they are applied.

你的浏览器遥控器。导航、点击、填表、抓数据、截图——说一句话就行。

使用方式

/auto-browser 打开 GitHub 看看我的 notifications
/auto-browser 去 Amazon 搜索 mechanical keyboard 截图前三个结果
/auto-browser 登录后台，把订单列表导出来
/auto-browser 帮我在这个页面点「下一步」然后填写地址表单

自然语言描述意图即可，不需要写代码。

必须使用 playwright-cdp 工具集

所有浏览器操作使用 user-playwright-cdp 的工具（连接真实 Chrome，保留登录态）。

• ✅ 用：user-playwright-cdp 的 browser_navigate、browser_snapshot、browser_click 等
• ❌ 禁止：cursor-ide-browser 的同名工具（沙盒浏览器，无登录态）

---

核心工作流

0. 确保 Chrome Debug 在线

curl -s http://127.0.0.1:9222/json/version

• ✅ 有响应 → 继续
• ❌ 无响应 → 直接启动，不问用户：

nohup /Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome \
  --remote-debugging-port=9222 \
  --user-data-dir="/tmp/chrome_debug_profile" \
  > /tmp/chrome_debug.log 2>&1 &

等 2 秒后重新 curl 确认。

1. 感知：snapshot 先行

每次操作前都先 browser_snapshot，获取页面无障碍树。这是你的「眼睛」。

• 用 snapshot 中的 ref 来定位元素
• 不确定页面状态时，先 snapshot 再决定下一步

2. 行动：选择合适的操作

根据用户意图选择操作，可自由组合：

意图	工具	说明
打开网页	browser_navigate	推断 URL，直接导航
后退	browser_navigate_back
点击按钮/链接	browser_click	用 snapshot 的 ref
填写输入框	browser_type（追加）/ browser_fill_form（清空后填）
选择下拉框	browser_select_option
上传文件	browser_file_upload	需绝对路径
按键	browser_press_key	Enter、Escape、Tab 等
悬停	browser_hover	展开菜单、tooltip
拖拽	browser_drag	startRef → endRef
处理弹窗	browser_handle_dialog	alert/confirm/prompt
等待加载	browser_wait_for	等时间或等文本出现/消失
执行 JS	browser_evaluate	页面没有暴露 UI 时的后备手段
管理标签页	browser_tabs	list/new/close/select
调整窗口	browser_resize	测试响应式布局

3. 确认：截图反馈

操作完成后 browser_take_screenshot 截图给用户确认结果。

---

操作原则

URL 推断

根据用户描述直接推断 URL 并导航，不要反问：

• 「打开 Google」→ https://www.google.com
• 「去 GitHub」→ https://github.com
• 「看看 V2EX」→ https://www.v2ex.com
• 模糊描述 → 用常识判断，导航后截图确认

只有完全无法推断时才问。

多步操作

复杂任务拆成步骤，每步遵循：snapshot → 操作 → 等待 → snapshot 确认。

示例：「登录后台导出订单」
1. navigate 到登录页 → snapshot
2. 如果已登录跳过，否则填写账号密码 → click 登录
3. wait_for 页面加载 → snapshot
4. click 导航到订单页 → snapshot
5. click 导出按钮 → 截图确认

等待策略

页面变化后（导航、点击、提交），用短间隔等待 + snapshot 确认：

browser_wait_for time=2 → snapshot → 检查是否就绪
→ 没好？再 wait_for time=2 → snapshot

不要一次等太久。2-3 秒一轮，最多重试 3 次。

数据提取

需要从页面抓取数据时：

1. browser_snapshot 获取页面结构
2. 从无障碍树中提取所需信息
3. 信息不够时用 browser_evaluate 执行 JS 提取
4. 结果太长写入文件，回复给摘要 + 路径

表单填写

填表场景遵循 auto-fill 的规则：

• 语义匹配字段，不要求精确
• 不确定的字段列出来问用户
• 密码类字段填前确认
• 多字段优先用 browser_fill_form 批量填写
• 填完截图确认

---

安全边界

操作	Agent 直接做	需用户确认
导航、浏览、截图	✅
点击普通按钮/链接	✅
填写非敏感字段	✅
读取/提取页面数据	✅
填写密码		✅
点击「提交」「付款」「删除」		✅（除非用户明确说"帮我提交"）
发送消息/邮件		✅
关闭标签页		✅（除非用户要求）

原则：只读操作自由做，写入/不可逆操作先确认。

---

错误处理

问题	处理
元素找不到	重新 snapshot，ref 可能变了
页面加载慢	browser_wait_for 等待，不要盲目重试
弹窗阻断	browser_handle_dialog 处理
需要登录	告诉用户，等用户手动登录后继续
验证码/人机验证	截图告知用户，等用户处理
JS 报错	browser_console_messages 查看错误日志

---

环境配置（首次）

如果 ~/.cursor/mcp.json 里没有 playwright-cdp 配置，添加：

"playwright-cdp": {
  "command": "npx",
  "args": ["-y", "@playwright/mcp@latest", "--cdp-endpoint", "http://127.0.0.1:9222"]
}

添加后提示用户重载 MCP（Cursor 设置 → MCP → Reload）。