Stealth Browser (Anti-Bot Detection)

Standard headless browsers get blocked. Amazon returns CAPTCHAs. LinkedIn serves login walls. eBay shows blank pages. These sites fingerprint navigator.webdriver, check for missing Chrome runtime objects, and detect headless user agents.

The stealth browser extension wraps puppeteer-extra with the puppeteer-extra-plugin-stealth plugin to bypass these checks. It presents as a real browser session — complete with spoofed plugins, valid Chrome runtime, and rotated user agents.

Anti-Detection Features

The stealth plugin applies several evasion techniques automatically:

• navigator.webdriver removal — Headless Chrome sets navigator.webdriver = true. Stealth deletes it.
• Chrome runtime injection — Sites check for window.chrome.runtime. Stealth injects a convincing stub.
• User agent rotation — Each session gets a real-world desktop user agent string, not the default HeadlessChrome UA.
• Plugin and MIME type spoofing — navigator.plugins returns a realistic plugin list instead of an empty array.
• WebGL vendor masking — GPU renderer strings match real hardware profiles.
• Iframe contentWindow patching — Prevents cross-frame detection of automation.

Tools

The extension provides six tools:

Tool	Description
stealth_navigate	Navigate to a URL with full anti-detection
stealth_scrape	Extract page content (text, HTML, or structured data)
stealth_click	Click elements by CSS selector
stealth_type	Type text into form fields
stealth_screenshot	Capture a screenshot of the current page
stealth_snapshot	Get the full DOM snapshot for analysis

Enabling

wunderland extensions enable stealth-browser

No API keys required. The extension uses Chromium bundled with Puppeteer.

Or add it to your agent.config.json:

{
  "extensions": {
    "tools": ["stealth-browser"]
  }
}

When to Use Which Browser Tool

Wunderland ships three ways to fetch web content. Pick the right one:

Tool	Use When	Speed	Detection Risk
web_search	You need search results, not a specific page	Fast	None (uses search API)
browser_navigate	The target site doesn't block bots (docs, wikis, public APIs)	Medium	Moderate
stealth_navigate	The site actively blocks headless browsers (Amazon, eBay, LinkedIn, etc.)	Slower	Low

Start with browser_navigate. If it fails with a bot detection error or returns empty content, switch to stealth_navigate. The agent can learn this pattern automatically (see below).

Automatic Failure Recovery

When browser_navigate fails on a bot-protected site, two systems kick in:

ToolFailureLearner

The ToolFailureLearner detects anti-bot failures (empty responses, CAPTCHA markers, 403 status codes) and records a lesson into the agent's RAG memory:

"browser_navigate failed on amazon.com due to anti-bot detection.
Use stealth_navigate for this domain."

On future queries involving that domain, RAG surfaces the lesson and the agent uses stealth_navigate from the start.

Agent Auto-Suggestion

Even without stored lessons, the agent's system prompt includes fallback guidance: if browser_navigate returns an error mentioning "bot", "captcha", or "access denied", retry with stealth_navigate. No configuration needed.

Example Session

You: Scrape the price of the Sony WH-1000XM5 on Amazon

Agent: I'll check Amazon for that product.
       [uses stealth_navigate → amazon.com/dp/B0BX5...]
       [uses stealth_scrape → extracts price element]

       The Sony WH-1000XM5 is currently listed at $278.00 on Amazon.

Limitations

• JavaScript-heavy SPAs — Some sites load content via complex JS frameworks. The stealth browser waits for networkidle0 by default, but extremely dynamic pages may need explicit wait selectors.
• CAPTCHAs — Stealth avoids triggering most CAPTCHAs, but sites with aggressive bot detection (e.g., Cloudflare Turnstile on highest settings) may still challenge the browser.
• Rate limiting — Rapid sequential requests to the same domain can trigger IP-based blocks regardless of stealth. Space requests naturally in conversation.
• Session persistence — Each stealth_navigate call starts a fresh browser context. Cookies and login state do not persist across calls.