Browser Web Search Skill

把任何网站变成命令行 API。13 平台 41 命令 — 知乎、小红书、B站、GitHub、豆瓣等。专为 OpenClaw 设计，复用浏览器登录态。

Security Scan

The skill's requests and instructions are internally consistent with its purpose (turning websites into a CLI using the browser session), but it depends on an external npm package that will execute JavaScript inside logged-in pages — this is expected for the feature but high-risk and should be audited before use.

Purpose & Capability

Name/description, declared requirements, and runtime behavior align: the skill turns websites into CLI commands by driving an OpenClaw-controlled browser and therefore legitimately needs the ability to execute JS in page contexts. It does not request unrelated secrets or services. The only notable requirement is a globally installed npm CLI package (browser-web-search), which is consistent with providing a 'bws' command-line tool.

Instruction Scope

SKILL.md instructs installing and invoking the external 'bws' CLI which in turn calls OpenClaw's browser evaluate to run JS in authenticated pages. That behavior is within the stated purpose, and the skill explicitly documents that it will read DOM and session-bound API responses. The instructions do not themselves direct exfiltration or access to unrelated system files, but executing third-party code inside logged-in pages inherently allows access to sensitive page content (messages, profile, orders).

Install Mechanism

Installation is via npm (npmjs.com) and references a GitHub repo — a common and acceptable distribution method. Risk is moderate because the npm package is not bundled in the skill and would be globally installed (-g), allowing arbitrary code execution on the host. The SKILL.md/skill.json acknowledge this and recommend auditing the repo and npm metadata before install.

Credentials

The skill does not request unrelated environment variables, credentials, or config paths. The sensitive capability is browser-session access (via OpenClaw), which is necessary for the advertised functionality but is high sensitivity; that access is documented rather than silently requested via env vars.

Persistence & Privilege

The skill is not always-enabled and does not request elevated platform privileges. It may use an optional ~/.bws/ config path. Autonomous invocation (model-invoked runs) is allowed by default on the platform, which combined with page-session access increases potential impact — the SKILL.md notes the sensitivity. No instructions indicate modifying other skills or system-wide settings.

Scan Findings in Context

[no_code_files_detected] expected: The repository scanned by regex had no code files bundled with the skill (instruction-only). This is expected because SKILL.md documents that runtime behavior is provided by an external npm package.

[npm_install_command_present] expected: The SKILL.md/skill.json include an npm install -g browser-web-search command. This is consistent with providing a CLI tool (bws), but it means the host will run third-party code and should be audited prior to installation.

Assessment

This skill appears to do what it claims, but it gains high-sensitivity access by executing JavaScript inside your logged-in browser pages via OpenClaw. Before installing or using it: - Audit the upstream package and GitHub repo (review adapter scripts that run in page context). Ensure published npm package matches repository source. - Prefer not to install globally on your primary machine/profile. Instead test in an isolated environment (VM, container, or dedicated browser profile with no sensitive accounts). - Do not use it with accounts that contain sensitive data (banking, private messages) until you trust the code. - Check npm download counts, issue history, recent commits, and package maintainer reputation. - Consider installing locally (not -g) or pinning a reviewed package version, and avoid running unreviewed versions. - If you are concerned about autonomous agent invocation, require explicit user approval for actions that access sensitive sites or disable agent autonomy for this skill where possible. If you cannot or will not audit the external package, treat this skill as high-risk and avoid installing it in production or on machines/accounts with sensitive data.

Browser Web Search (BWS) Skill

把任何网站变成命令行 API，专为 OpenClaw 设计，复用浏览器登录态。

🏗️ 架构说明

OpenClaw/AI Agent
    ↓ (读取 Skill 配置)
browser-web-search-skill
    ↓ (调用 CLI)
bws 命令
    ↓ (OpenClaw Browser)
目标网站

🎯 项目特点

专为 OpenClaw 设计

• 零配置：无需 Chrome Extension、无需 Daemon，开箱即用
• 深度集成：直接使用 OpenClaw 浏览器，与其他 Skill 共享登录态
• 轻量精简：核心代码仅 22KB，无运行时依赖

复用登录态

• 无需 API Key：使用你在浏览器中的登录状态
• 绕过反爬：请求来自真实浏览器，不会被封禁
• 隐私安全：数据在本地处理，不经过第三方服务器

AI Agent 友好

• 结构化输出：所有命令返回 JSON，便于 AI 解析
• jq 过滤：内置 jq 支持，精确提取所需数据
• 错误提示：清晰的错误信息和修复建议

📋 前置要求

安装 browser-web-search

npm install -g browser-web-search

验证安装

bws --version
bws site list

🚀 快速开始

# 查看所有可用命令
bws site list

# 运行 adapter
bws zhihu/hot                      # 知乎热榜
bws xiaohongshu/search "旅行"       # 小红书搜索
bws bilibili/popular               # B站热门
bws github/repo facebook/react     # GitHub 仓库

📊 内置平台（13 平台 41 命令）

分类	平台	命令数	示例命令
搜索	Google, Baidu, Bing	3	bws google/search "query"
社交	小红书, 知乎	10	bws zhihu/hot
新闻	36kr, 今日头条	3	bws 36kr/newsflash
开发	GitHub, CSDN, 博客园	8	bws github/repo owner/repo
视频	Bilibili	9	bws bilibili/popular
娱乐	豆瓣	6	bws douban/top250
招聘	BOSS直聘	2	bws boss/search "职位"

🔧 标准操作流程 (SOP)

操作 1：查看可用命令

场景：用户想知道有哪些可用的 adapter

命令：

bws site list

输出示例：

zhihu/
  hot                  - Get Zhihu hot list
  search               - Search Zhihu
  question             - Get question details
  me                   - Get logged-in user info

xiaohongshu/
  search               - 搜索小红书笔记
  note                 - 获取笔记详情
  ...

---

操作 2：搜索 adapter

场景：用户想找特定平台的命令

命令：

bws site search bilibili

输出示例：

bilibili/popular       Get Bilibili popular videos
bilibili/search        Search Bilibili videos
bilibili/video         Get video details
...

---

操作 3：查看 adapter 详情

场景：用户想了解某个命令的参数

命令：

bws site info bilibili/video

输出示例：

bilibili/video - Get Bilibili video details by bvid

参数:
  bvid (required)      视频 BV 号

示例:
  bws site bilibili/video BV1xx411c7mD

---

操作 4：获取知乎热榜

场景：用户想获取知乎热门话题

命令：

bws zhihu/hot

输出示例：

{
  "items": [
    {
      "title": "如何评价...",
      "url": "https://www.zhihu.com/question/...",
      "heat": "1234万热度"
    }
  ]
}

---

操作 5：搜索小红书

场景：用户想搜索小红书内容

命令：

bws xiaohongshu/search "旅行攻略"

输出示例：

{
  "notes": [
    {
      "id": "abc123",
      "title": "云南旅行攻略",
      "author": "旅行博主",
      "likes": 1234
    }
  ]
}

---

操作 6：获取 B站热门视频

场景：用户想看 B站热门

命令：

bws bilibili/popular

输出示例：

{
  "videos": [
    {
      "bvid": "BV1xx411c7mD",
      "title": "视频标题",
      "author": "UP主",
      "play": "100万",
      "like": "5万"
    }
  ]
}

---

操作 7：使用 jq 过滤

场景：用户只需要部分数据

命令：

# 只获取标题
bws zhihu/hot --jq '.items[].title'

# 提取特定字段
bws bilibili/popular --jq '.videos[] | {title, play}'

---

操作 8：获取 GitHub 仓库信息

场景：用户想查看某个 GitHub 仓库

命令：

bws github/repo facebook/react

输出示例：

{
  "name": "react",
  "description": "A declarative, efficient, and flexible JavaScript library...",
  "stars": 220000,
  "forks": 45000,
  "language": "JavaScript"
}

---

操作 9：搜索引擎搜索

场景：用户想使用搜索引擎

命令：

bws google/search "OpenClaw AI"
bws baidu/search "人工智能"
bws bing/search "machine learning"

---

🔧 技术架构：如何访问登录态

BWS 不直接读取浏览器 Cookie 文件或用户配置文件。它通过 OpenClaw 提供的 API 与浏览器交互：

bws 命令
    ↓ 调用
openclaw browser evaluate <script>
    ↓ 在已打开的标签页中执行 JavaScript
目标网站（使用该标签页的登录态）

工作原理

1. BWS 调用 OpenClaw CLI：

   openclaw browser evaluate --domain "zhihu.com" "<adapter-script>"
   

2. OpenClaw 在浏览器标签页中执行脚本：

   • 找到匹配域名的已打开标签页
   • 或打开新标签页访问目标网站
   • 在页面上下文中执行 adapter 脚本

3. 脚本在页面中运行：

   • 脚本以网页的身份运行（如同 DevTools Console）
   • 自动继承该页面的登录态（Cookie、Session）
   • 通过 DOM 操作或 fetch 获取数据

数据访问范围

访问内容	是否访问	说明
浏览器 Cookie 文件	❌ 否	不直接读取 ~/.config/chromium/Cookies 等文件
用户配置目录	❌ 否	不访问 ~/.bws/ 以外的配置
其他网站数据	❌ 否	只能访问 adapter 指定的域名
当前页面 DOM	✅ 是	adapter 脚本在页面中执行
当前页面 Session	✅ 是	继承页面的登录状态

安全边界

• 隔离性：每个 adapter 只能访问其声明的 domain
• 透明性：所有 adapter 代码是公开的 JS 文件，可审计
• 无持久化：BWS 不保存任何 Cookie 或 Session Token
• 用户控制：登录操作由用户在浏览器中手动完成

---

⚠️ 登录态管理

如果网站需要登录，命令会返回 401/403 错误。

解决步骤：

1. 在 OpenClaw 浏览器中打开网站：

   openclaw browser open https://xiaohongshu.com
   

2. 手动完成登录（BWS 不参与此过程）

3. 重试命令：

   bws xiaohongshu/me
   

注意：BWS 只是在已登录的页面中执行脚本，不会存储或传输你的登录凭证。

---

📝 输出格式

所有命令默认返回 JSON 格式：

成功响应：

{
  "items": [...],
  "count": 10
}

错误响应：

{
  "success": false,
  "error": "错误信息"
}

---

🎓 示例对话

用户：帮我看看知乎今天有什么热门话题

AI：好的，我来获取知乎热榜。

bws zhihu/hot

AI：以下是知乎热榜前 10：

1. 如何评价...（1234万热度）
2. 为什么...（890万热度） ...

---

用户：搜索一下小红书上关于"露营"的笔记

AI：好的，我来搜索小红书。

bws xiaohongshu/search "露营"

AI：找到以下相关笔记：

1. 《新手露营装备清单》- 点赞 5.2k
2. 《周末露营好去处》- 点赞 3.8k ...

---

🔒 安全性说明

• ✅ 所有操作在本地执行
• ✅ 使用 OpenClaw 浏览器的登录态
• ❌ 不会收集用户信息
• ❌ 不会上传到第三方服务器

---

📚 参考资料

• 项目 GitHub
• browser-web-search 核心库
• npm 包

---

📝 维护说明

• 版本: 0.1.0
• 最后更新: 2026-03-31
• 维护者: Ping Si [e-mail]
• 许可证: MIT

---

✅ 首次成功检查清单

新用户应该能在 2 分钟内完成：

• 安装工具：npm install -g browser-web-search
• 检查安装：bws --version
• 查看命令：bws site list
• 测试运行：bws zhihu/hot
• 看到 JSON 输出

如果以上步骤都能顺利完成，说明 Skill 已正确配置！