Agent Browser Cli.Bak

使用 agent-browser CLI 进行浏览器自动化。用于签到、填表、截图、信息抓取等需要控制浏览器的任务。触发条件：(1) 用户要求自动化浏览器操作 (2) 需要签到、填表、点击按钮 (3) 需要抓取网页内容作为研究素材

Security Scan

The skill's instructions match a browser-automation purpose, but metadata inconsistencies and an instruction to install an unverified npm package introduce non-trivial risks the user should review before installing.

Purpose & Capability

SKILL.md describes a browser-automation CLI (open, click, fill, snapshot, screenshot) that matches the skill name and description. However, the package is presented as a Vercel product in the text while source/homepage are unknown, and registry metadata (owner/slug) does not match the included _meta.json; these inconsistencies reduce confidence in provenance.

Instruction Scope

Runtime instructions are narrowly scoped to browser automation tasks (open, click, fill, snapshot, screenshot) and examples reference only local scripts and files (e.g., ~/.openclaw/scripts, /tmp screenshots). There is no instruction to read unrelated system files or exfiltrate data in the SKILL.md itself.

Install Mechanism

The SKILL.md tells users to run `npm install -g agent-browser` and `agent-browser install`. Installing an unverified global npm package can execute arbitrary code (postinstall scripts), and `agent-browser install` likely downloads browser engines. Because the skill is instruction-only, there is no bundled code to inspect; the install step is the primary risk here.

Credentials

The skill declares no required environment variables, credentials, or config paths. Example usage shows filling a password as a demonstration only. There is no explicit request for unrelated secrets or multiple credentials.

Persistence & Privilege

Flags show no forced persistence (always:false) and normal autonomous invocation behavior. The skill suggests creating a user script under ~/.openclaw/scripts, which is reasonable for automation tasks and does not request system-wide privileges.

What to consider before installing

This skill appears to be a straightforward browser-automation instruction set, but two things warrant caution before installing or running anything: 1) Provenance mismatch: the package/source is unknown, the SKILL.md claims Vercel, and the included _meta.json has different owner/slug values than the registry metadata—this could indicate copying or tampering. Verify the package author and homepage on the npm registry or GitHub before trusting it. 2) npm install risk: the instructions ask you to run `npm install -g agent-browser`. Global npm installs and package postinstall scripts can execute arbitrary code and may download additional binaries (browser engines). Inspect the package contents and repository (or prefer a well-known vendor package). If you must try it, run in a disposable environment (container, VM) or sandbox, review postinstall scripts, and avoid storing real credentials in automation scripts. Additional checks that would increase confidence: official homepage or GitHub repo matching the claimed owner, consistent registry metadata, a review of the npm package source (no unexpected network endpoints or postinstall scripts), and confirmation that `agent-browser install` only downloads expected browser engines.

Agent Browser

Vercel 出品的浏览器自动化 CLI，基于 Playwright，比标准浏览器工具更快更灵活。

快速开始

agent-browser open <url>     # 打开网页
agent-browser snapshot       # 获取页面可访问性树
agent-browser click @<ref>   # 点击元素（用ref引用）
agent-browser fill @<ref> "内容"  # 填入内容
agent-browser close         # 关闭浏览器

常用命令

导航

agent-browser open <url>      # 打开URL（别名：goto, navigate）
agent-browser back            # 后退
agent-browser forward         # 前进
agent-browser reload          # 刷新

交互

agent-browser click <sel>                    # 点击
agent-browser dblclick <sel>                  # 双击
agent-browser fill <sel> "text"               # 填入（清空后填）
agent-browser type <sel> "text"               # 输入（追加）
agent-browser select <sel> <value>             # 选择下拉选项
agent-browser check <sel>                      # 勾选复选框
agent-browser uncheck <sel>                   # 取消勾选
agent-browser press <key>                      # 按键（Enter, Tab, Escape等）

获取信息

agent-browser snapshot              # 获取可访问性树（推荐）
agent-browser get text <sel>        # 获取文本
agent-browser get html <sel>        # 获取HTML
agent-browser get value <sel>       # 获取输入值
agent-browser get title             # 获取页面标题
agent-browser get url               # 获取当前URL
agent-browser screenshot [path]     # 截图
agent-browser screenshot --annotate  # 带标注的截图

元素定位

通过 snapshot 输出的 ref（如 @e14）直接引用：

agent-browser click @e14
agent-browser fill @e13 "hello"

或使用 CSS 选择器：

agent-browser click "#submit"
agent-browser fill "input[name='email']" "[e-mail]"

或使用 ARIA 角色查找：

agent-browser find role button click --name "Submit"
agent-browser find text "Sign In" click
agent-browser find label "Email" fill "[e-mail]"
agent-browser find placeholder "Search" type "query"

典型工作流

1. 签到任务

# 打开登录页
agent-browser open <签到页面URL>

# 获取页面结构
agent-browser snapshot

# 点击登录/签到按钮（用实际ref替换 @eXX）
agent-browser click @eXX

# 等待页面加载
sleep 2
agent-browser snapshot

2. 填表任务

agent-browser open <表单URL>
agent-browser snapshot

# 填入各字段
agent-browser find label "用户名" fill "myuser"
agent-browser find label "密码" fill "mypassword"
agent-browser find role button click --name "提交"

3. 定时签到（配合cron）

创建脚本 ~/.openclaw/scripts/daily-checkin.sh：

#!/bin/bash
agent-browser open <签到URL>
sleep 2
agent-browser find role button click --name "签到"
agent-browser screenshot /tmp/checkin_$(date +%Y%m%d).png
agent-browser close

注意事项

1. 先 snapshot 再操作 - 每次页面变化后重新获取 ref
2. 添加等待 - 页面加载需要时间，用 sleep 2 或等待
3. 保持浏览器开启 - 多个操作可以在同一浏览器会话中完成
4. 完成后关闭 - 用 agent-browser close 释放资源

依赖安装

如果 agent-browser 未安装：

npm install -g agent-browser
agent-browser install