web-browsing（网页浏览）

Web Browsing Skill

Automatic fallback mechanism for web research.

CRITICAL: Tool Availability

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)

Fallback Chain

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

Step 1: WebFetch (Default)

WebFetch url="https://example.com"

Continue to Step 2 if:

• HTTP error (4xx, 5xx)
• Content empty or truncated
• Authentication required
• Rate limited

Step 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

Step 3: User Assistance

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.

Detection Patterns

Captcha/Block Indicators

• "verify you are human"
• "captcha"
• "cloudflare"
• "access denied"
• "please wait while we verify"
• Blank page with only scripts
• HTTP 403, 429, 503

Success Indicators

• Meaningful content returned
• Expected HTML structure
• No error messages

Best Practices

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

Integration with ai_interaction

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