Browser Ops Publish

AI Agent 的网页访问路由决策指南。全 CLI 架构，零 MCP 依赖，不占常驻上下文 token。 按成本逐级升级: WebFetch($0) → opencli web read($0,带Cookie) → Firecrawl → agent-browser → browser-use。 IMPORTA...

Security Scan

The skill's stated metadata omits several sensitive requirements and persistent behaviors (cookie/state storage, API keys, proxy credentials, Chrome extension access) that its SKILL.md and scripts clearly rely on — the components are coherent for a web-crawling/router tool, but the declared surface (no env, no config paths) is inconsistent and deserves caution.

Purpose & Capability

The name/description claim a pure CLI routing guide with 'no required env vars' and 'no config paths', but the README, SKILL.md and scripts require or strongly recommend many external tools and secrets: opencli (Chrome extension/daemon), agent-browser, browser-use, Tavily/FIRECRAWL API keys, Zendriver/Camoufox, proxy credentials (PROXY_URL), and local cookie/state files. Some of these (Chrome extension access, cookie stores, proxy credentials) are sensitive; the registry metadata does not declare them.

Instruction Scope

Runtime instructions explicitly direct automated escalation (e.g., 'on WebFetch → 403 automatically upgrade to opencli web read — DON'T ASK USER') and reuse of Chrome cookies via an extension, plus scripts for exporting/importing and injecting cookies into Playwright contexts (~/.browser-ops/unified-state.json). That means the agent will access and persist SSO/cookie tokens and can inject them into headless browsers. The instructions also advise anti-detection and proxy use for bypassing Cloudflare and fingerprinting — broad scope that touches sensitive state and network configuration.

Install Mechanism

The skill is instruction-only (no automatic install spec), which reduces automatic risk. However the repository includes many helper scripts that assume installing npm packages and pip packages (opencli, agent-browser, tavily-python, firecrawl, zendriver, camoufox). Those installs are standard package managers (npm/pip) and not direct-download URLs, but will pull third-party code into the environment if the user follows the setup instructions.

Credentials

Registry metadata declares no required env vars, but the content references and expects multiple environment values (TAVILY_API_KEY, FIRECRAWL_API_KEY, PROXY_URL, HTTP_PROXY/HTTPS_PROXY/SOCKS_PROXY). The skill also documents storing proxy credentials in PROXY_URL and persisting cookie state on disk. Requesting or using such secrets is proportionate to advanced crawling features, but the metadata omission is misleading and increases risk because users may not realize sensitive credentials or cookie tokens will be reused or stored.

Persistence & Privilege

Scripts create and use a persistent cookie store at ~/.browser-ops/cookie-store/unified-state.json and profile dirs; sync-cookies.sh can export/import and inject cookies into Playwright contexts. While the skill does not set 'always:true', it does instruct persistent state creation and reuse of SSO cookies without explicit per-use prompts (SKILL.md instructs automatic upgrades 'don't ask user' in some cases). Persisting SSO/cookie tokens and automatically reusing them increases sensitive privilege and potential for misuse or accidental leakage if the store is not secured.

What to consider before installing

This skill appears to implement a coherent web-access routing policy, but there are important surprises and risks you should consider before installing or running it: - Mismatch between declared metadata and actual behavior: the registry claims no env vars/config paths, but the SKILL.md and scripts rely on API keys (TAVILY_API_KEY, FIRECRAWL_API_KEY), proxy credentials (PROXY_URL/HTTP_PROXY), and a Chrome Extension (opencli) and will create ~/.browser-ops cookie/state files. Treat those as required for full functionality. - Sensitive state is persisted: sync-cookies.sh saves a unified-state.json in your home directory containing cookies and SSO tokens. That file can be injected into automated browsers. If you run these scripts, inspect and protect ~/.browser-ops/cookie-store/unified-state.json (chmod 600) and don't store it on shared machines. - Automatic cookie reuse and 'no-user-prompt' escalation: the skill instructs agents to automatically upgrade to tools that reuse browser cookies (opencli web read) when WebFetch 403s, and explicitly says 'do not ask the user'. If you expect explicit permission before reusing logged-in sessions, this behavior is a privacy/safety concern. - External credentials and proxies: the anti-detection guidance recommends residential proxy providers and passing credentials via PROXY_URL. Only provide such credentials in secure environments and avoid embedding them in command-line arguments or public logs. - Install and run in an isolated environment if possible: if you decide to try this skill, run it on a VM or container where potential leakage of cookies or proxy credentials is acceptable. Review the scripts (sync-cookies.sh, web-read.sh, web-search.sh, anti-detection code) before executing them. - Verification steps: check for ~/.browser-ops files after any run; audit unified-state.json contents for sensitive cookies; ensure you understand which third-party packages you will install (npm/pip) and their trustworthiness. - If you want to proceed but minimize risk: (1) do not run sync-cookies.sh export/login unless you intend to persist cookies; (2) avoid importing cookies into automated contexts; (3) do not set PROXY_URL with credentials in a global shell session; (4) prefer ephemeral Docker containers and isolated user accounts. Given the clear inconsistencies between declared metadata and actual required secrets/state, I recommend treating this skill as 'suspicious' until the author updates the metadata to declare required env vars and config paths and clarifies the automatic cookie reuse/consent behavior.

Browser Operations — 网页访问路由决策指南

全 CLI 架构。所有工具通过 Bash 调用，零 MCP 依赖，不占常驻上下文。 核心原则：能用 HTTP 就不开浏览器，能用 opencli 就不用 browser-use。

核心规则

MUST 从免费层开始。NEVER 直接跳到浏览器工具。

每次网页任务按这个顺序判断，命中就停： 0. 目标是 opencli 已适配平台？→ 直接 opencli <platform> <command>，跳过 WebFetch 判断方法：URL 域名或用户意图命中已适配站点（opencli list 查看完整列表，74 个站点 454 条命令）。 常见已适配平台：twitter/x.com, zhihu, weibo, bilibili, hackernews, xiaohongshu, douyin, reddit, youtube, github, arxiv, bloomberg, linkedin, douban, jd, wikipedia, v2ex, tieba, spotify, steam, medium, substack, producthunt, stackoverflow, 36kr, weread, xueqiu 等。 这些平台有专属适配器，返回结构化数据，比 web read 更准确更高效。

1. WebFetch/WebSearch 能搞定（非已适配平台）？→ 用它，$0
2. 403/SSO/空？→ opencli web read，$0，Cookie 零配置
3. 需要 JS 渲染/结构化？→ firecrawl scrape "url"
4. 需要交互（≤3 步）？→ opencli operate，Cookie 零配置
5. 需要精确控制/Ref/录制？→ agent-browser，@e1 稳定引用
6. 需要 AI 自主多步？→ browser-use -p "任务"，$0.01-0.05/步
7. 被反爬拦截？→ zendriver

违反顺序 = 浪费钱或丢失数据。已适配平台用 opencli web read 只能拿到 HTML，用 opencli <platform> 能拿到结构化字段。

<anti-example> 用户: "搜一下 Twitter 上关于 AI agent 的讨论" 错误: opencli web read --url "https://x.com/search?q=AI+agent" → 拿到 HTML 残片 正确: opencli twitter search "AI agent" → 结构化推文列表 (id, author, text, likes, views) </anti-example> <anti-example> 用户: "帮我看看知乎热榜" 错误: WebFetch("https://www.zhihu.com/hot") → 空/SPA 空壳 正确: opencli zhihu hot → 结构化热榜列表 </anti-example> <anti-example> 用户: "帮我看看 https://example.com/article 的内容" 错误: browser-use -p "extract content from example.com" → $0.05 正确: WebFetch("https://example.com/article") → $0 (example.com 不是已适配平台) </anti-example> <anti-example> 用户: "帮我读一下内部网站的这篇文章 https://internal.company.com/doc/123" 错误: WebFetch → 403 → 放弃，说"无法访问" 正确: WebFetch → 403 → 自动升级到 opencli web read (带 Chrome Cookie) </anti-example>

条件判断规则

当 WebFetch 返回 403 或 302 到登录页时 → 自动升级到 opencli web read，不要询问用户。 当 opencli web read 返回 exit 77 时 → SSO/Cookie 过期，MUST 停止降级，直接提示用户回 Chrome 登录后重试。后续工具 (firecrawl/agent-browser) 也没有登录态，继续降级只是浪费。 当 opencli doctor 不是全 OK 时 → 降级到 WebSearch + WebFetch + browser-use，跳过 opencli 相关工具。 当任务需要填表但不确定表单字段时 → 先用 opencli operate state 获取可交互元素列表，确认字段后再操作。 如果不确定用哪个工具 → 询问用户或从最便宜的 WebFetch 开始逐级升级。

已适配平台 MUST 直接用 opencli <platform>，NEVER 走 WebFetch/web read 通路，otherwise 丢失结构化数据。 非已适配平台 MUST 从 WebFetch 开始，otherwise 每个请求多花 $0.01-0.05 且速度慢 10 倍。 NEVER 用 browser-use 做简单读取，otherwise 一次 $0.05 的操作 WebFetch $0 就能完成。 不要跳过 opencli web read 直接用 Firecrawl，而是先试免费的 Cookie 路径。

Output

返回给用户时 MUST 包含:

• 提取到的内容（Markdown 格式）
• 实际使用的工具和原因（如 "WebFetch 返回 403，已自动升级到 opencli web read"）
• 如果全部失败，给出具体原因和用户可操作的建议（如 "需要在 Chrome 中重新登录"）

returns: 网页内容 (Markdown) + 工具链路径 + 失败原因（如有）

路由决策树

收到任务
│
├─ 0. opencli 可用? → opencli doctor (3 个 OK)
│     否 → 降级: WebSearch + WebFetch + browser-use
│
├─ 1. 目标是已适配平台? (URL 域名 或 用户意图 命中 opencli list 中的站点)
│     是 → 直接 opencli <platform> <command>，跳过 WebFetch/web read
│     例: twitter/x.com → opencli twitter search/timeline/trending
│         zhihu.com → opencli zhihu hot/search/answer
│         bilibili.com → opencli bilibili search/hot/subtitle
│
├─ 2. 要搜索（没 URL，非已适配平台）?
│     ├─ WebSearch (内置, 始终可用)
│     ├─ 深度搜索 → tavily search "query" --search-depth advanced
│     ├─ 独立索引 → curl brave API
│     └─ fallback → opencli google search
│
├─ 3. 有 URL，非已适配平台?
│     ├─ WebFetch ($0) → 403/SSO? → opencli web read ($0, Cookie 直连)
│     │     ├─ exit 77 (SSO 过期) → 停止降级，提示用户回 Chrome 登录
│     │     └─ 其他失败 (JS 不足/内容太短) → 继续降级到 firecrawl
│     └─ JS 渲染/PDF/结构化 → firecrawl scrape "url"
│
├─ 4. 要交互?
│     ├─ ≤3 步 → opencli operate (Cookie 零配置, 17 个命令)
│     │   open → state → click/type/select/scroll → screenshot → close
│     ├─ 需要稳定引用/录制/标注截图 → agent-browser (60+ 命令)
│     │   open → snapshot -i → click @e1 → screenshot --annotate
│     ├─ AI 自主多步 → browser-use -p "自然语言任务"
│     └─ 未知 DOM → Stagehand act("点击登录")
│
├─ 5. 被反爬? → python -c "import zendriver as zd; ..."
│
└─ 全失败 → 告知用户具体原因和建议，NEVER 静默失败

升级信号

当前工具返回	升级到	命令
WebFetch → 403/302 登录页	opencli web read	opencli web read --url "url"
WebFetch → 空/SPA 空壳	Firecrawl	firecrawl scrape "url"
opencli → exit 77	停止降级，提示用户	SSO 过期，后续工具也没登录态，直接提示用户回 Chrome 登录
需要点击/填表	opencli operate	opencli operate open "url" && state
编号 [N] 不稳定	agent-browser	agent-browser snapshot -i → click @e1
多步复杂任务	browser-use	browser-use -p "任务描述"
Cloudflare 拦截页	Zendriver	python3 -c "import zendriver..."

搜索工具

# 内置 (始终可用)
WebSearch → Claude Code 内置，直接调用

# Tavily — AI 原生搜索，返回 answer + results (免费 1000 次/月)
tavily search "query"                              # pip install tavily-python
tavily search "query" --search-depth advanced       # 深度模式
tavily extract "https://url"                        # URL 内容提取

# Brave — 独立索引，不依赖 Google/Bing
curl -s "https://api.search.brave.com/res/v1/web/search?q=query" \
  -H "X-Subscription-Token: $BRAVE_API_KEY"

# Firecrawl — JS 渲染 + Markdown 提取 (免费 500 次)
firecrawl scrape "https://url"                     # pip install firecrawl
firecrawl crawl "https://url" --limit 10           # 批量
firecrawl map "https://url"                        # URL 发现

# 平台搜索 — 75 站点结构化数据
opencli twitter trending / zhihu hot / hackernews top / xiaohongshu search "旅行"
opencli list                                       # 查看所有适配器

浏览器交互工具

# opencli operate — 简单交互，Cookie 零配置 (通过 Chrome Extension 直连)
opencli operate open "url"
opencli operate state                              # 可交互元素 [1][2][3]
opencli operate click 5 / type 3 "hello" / select 2 "选项A"
opencli operate scroll down / keys Enter / wait text "Success"
opencli operate screenshot /tmp/shot.png / network / close

# agent-browser — 复杂交互，@e1 Ref 引用 (基于 Accessibility Tree，页面重渲染不变)
agent-browser open "url" && agent-browser snapshot -i
agent-browser click @e2 / fill @e3 "hello"         # Ref 引用
agent-browser screenshot --annotate /tmp/a.png     # 标注截图
agent-browser record start                        # 录制操作
agent-browser batch "click @e1 && wait 1000 && screenshot"
agent-browser --auto-connect snapshot              # 连接已运行 Chrome
# 搭配 Lightpanda 省 token: snapshot ~500 token vs Chrome ~3000 token
# ./lightpanda serve --port 9222 && agent-browser connect 9222

# browser-use — AI 自主操作 (LLM 决策循环，每步 $0.01-0.05)
browser-use -p "去 example.com 注册账号"            # 自然语言驱动
browser-use --connect -p "任务"                     # 连接已运行 Chrome
browser-use run --remote "任务"                     # 云端执行

opencli operate vs agent-browser

维度	opencli operate	agent-browser
场景	≤3 步简单操作	复杂/精确操作
元素定位	[N] 编号（页面变了会变）	@e1 Ref（稳定，基于 ARIA role+name）
Cookie	Chrome Extension 直连，零配置	--profile / --auto-connect
标注截图	❌	✅ --annotate（给视觉模型用）
录制回放	❌	✅ record（固定流程自动化）
批量/DOM diff	❌	✅ batch / diff
内核替换	❌	✅ Lightpanda（省 80% token）
命令数	17	60+
iOS 测试	❌	✅ -p ios

前置条件

# 必装
npm i -g @jackwener/opencli
# Chrome 装 OpenCLI Browser Bridge: $(npm root -g)/@jackwener/opencli/extension
opencli doctor  # 3 个 OK 才能用

# 按需 (全 CLI，不需要配 MCP)
npm i -g agent-browser               # Ref 引用/录制/标注截图

> See references/ for extended content.