Back to Details

web-browsing

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Web Browsing Skill

网页浏览Skill

Automatic fallback mechanism for web research.
用于网页调研的自动降级机制。

CRITICAL: Tool Availability

重要提示：工具可用性

NEVER check mcps folder to verify tool existence. Tools are ALWAYS available - just call them directly:
  • WebFetch
    - Cursor built-in fetch tool (fastest)
  • cursor-browser-extension
    - Cursor built-in browser
  • user-chrome-devtools
    - Chrome DevTools (alternative)
绝对不要检查mcps文件夹来验证工具是否存在。 工具始终可用——直接调用即可：
  • WebFetch
    - Cursor内置的抓取工具（速度最快）
  • cursor-browser-extension
    - Cursor内置浏览器
  • user-chrome-devtools
    - Chrome开发者工具（备选方案）

Fallback Chain

降级流程链

1. WebFetch (fast, preferred)
   ↓ fails/error
2. Browser MCP (cursor-browser-extension)
   ↓ captcha/block
3. ai_interaction (request user help)
1. WebFetch (fast, preferred)
   ↓ fails/error
2. Browser MCP (cursor-browser-extension)
   ↓ captcha/block
3. ai_interaction (request user help)

Step 1: WebFetch (Default)

步骤1：WebFetch（默认选项）

WebFetch url="https://example.com"
Continue to Step 2 if:
  • HTTP error (4xx, 5xx)
  • Content empty or truncated
  • Authentication required
  • Rate limited
WebFetch url="https://example.com"
出现以下情况则进入步骤2：
  • HTTP错误（4xx、5xx状态码）
  • 内容为空或被截断
  • 需要身份验证
  • 触发速率限制

Step 2: Browser MCP

步骤2：Browser MCP

Use 
cursor-browser-extension
MCP tools directly:
1. browser_navigate → target URL
2. browser_snapshot → get page content
3. browser_take_screenshot → visual reference (if needed)
Alternative: 
user-chrome-devtools
if cursor-browser-extension fails.
Continue to Step 3 if:
  • Captcha detected
  • Bot check / "verify you're human"
  • Login required
  • Cloudflare/DDoS protection
  • Page content not loading
直接使用
cursor-browser-extension
MCP工具：
1. browser_navigate → target URL
2. browser_snapshot → get page content
3. browser_take_screenshot → visual reference (if needed)
备选方案： 如果cursor-browser-extension失败，使用
user-chrome-devtools
。
出现以下情况则进入步骤3：
  • 检测到验证码
  • 机器人验证 / "验证你是人类"
  • 需要登录
  • Cloudflare/DDoS防护
  • 页面内容无法加载

Step 3: User Assistance

步骤3：寻求用户协助

CRITICAL RULES:
  • NEVER make autonomous decisions (e.g., switching to DuckDuckGo when Google blocked)
  • NEVER bypass user by trying alternative search engines/sites without asking
  • ALWAYS inform user and wait for instruction via ai_interaction
Call 
ai_interaction
tool with message:
"Browser bị chặn bởi [captcha/bot check/login].
URL: [url]
Cần bạn:
1. Mở browser manual
2. Bypass protection
3. Báo lại khi xong để tôi tiếp tục

Hoặc bạn muốn thử [alternative option]?"
Then wait for user confirmation via ai_interaction before retrying.
重要规则：
  • 绝对不要自主做决策（例如，当Google被屏蔽时切换到DuckDuckGo）
  • 绝对不要在未询问的情况下尝试替代搜索引擎/网站来绕过用户
  • 必须告知用户并等待通过ai_interaction发来的指令
调用
ai_interaction
工具并发送以下消息：
"浏览器被[captcha/机器人验证/登录]拦截。
URL: [url]
需要您：
1. 手动打开浏览器
2. 绕过防护验证
3. 完成后告知我以继续

或者您想尝试[替代方案]？"
然后在重试前等待用户通过ai_interaction发来的确认。

Detection Patterns

检测模式

Captcha/Block Indicators

验证码/拦截标识

  • "verify you are human"
  • "captcha"
  • "cloudflare"
  • "access denied"
  • "please wait while we verify"
  • Blank page with only scripts
  • HTTP 403, 429, 503
  • "verify you are human"
  • "captcha"
  • "cloudflare"
  • "访问被拒绝"
  • "请等待我们验证"
  • 仅包含脚本的空白页面
  • HTTP 403, 429, 503

Success Indicators

成功标识

  • Meaningful content returned
  • Expected HTML structure
  • No error messages
  • 返回有意义的内容
  • 符合预期的HTML结构
  • 无错误消息

Best Practices

最佳实践

  1. Always try WebFetch first - fastest option
  2. Check response quality - not just HTTP status
  3. Don't spam retries - wait between attempts
  4. Be transparent - tell user what's happening
  5. Save useful content - don't re-fetch unnecessarily
  1. 始终优先尝试WebFetch - 速度最快的选项
  2. 检查响应质量 - 不要只看HTTP状态码
  3. 不要频繁重试 - 尝试之间要等待
  4. 保持透明 - 告知用户当前情况
  5. 保存有用内容 - 不要重复抓取不必要的内容

Integration with ai_interaction

与ai_interaction的集成

When user helps bypass protection:
  1. User completes manual action
  2. User responds via ai_interaction
  3. Agent retries with Browser MCP (cookies preserved)
  4. Continue with research
当用户协助绕过防护后：
  1. 用户完成手动操作
  2. 用户通过ai_interaction回复
  3. Agent使用Browser MCP重试（保留Cookie）
  4. 继续调研