shield_lock
JinDaGe
← Back to skills
extension
Category: Development & EngineeringNo API key required

stealthy-auto-browse

Browser automation that passes CreepJS, BrowserScan, Pixelscan, and Cloudflare — zero CDP exposure, OS-level input, persistent fingerprints. Use when standard browser skills get 403s or CAPTCHAs.

personAuthor: jakexiaohubgithub
open_in_newRepositorydownloadDownload package

stealthy-auto-browse

Why Not Just Use agent-browser?

Standard browser skills (agent-browser, etc.) use Playwright + Chromium which exposes Chrome DevTools Protocol signals. Bot detection services like Cloudflare, DataDome, PerimeterX, and Akamai detect these instantly. This skill uses Camoufox (custom Firefox) with zero CDP exposure and real OS-level mouse/keyboard input — completely invisible to behavioral analysis.

When To Use This Skill

  • Target site has bot detection (Cloudflare, DataDome, PerimeterX, Akamai)
  • Target site blocks headless browsers or returns CAPTCHAs
  • You need to maintain a logged-in session without getting banned
  • Standard browser skill is getting 403s or blocked responses
  • You need to scrape data from sites with anti-bot protection

When NOT To Use This Skill

  • Simple page fetches with no bot protection (use curl/WebFetch instead)
  • Sites that don't care about automation (use agent-browser, it's faster)
  • You only need to read static HTML (use curl)

Setup

1. Run the container:

docker run -d -p 8080:8080 -p 5900:5900 psyb0t/stealthy-auto-browse

2. Configure environment:

export STEALTHY_AUTO_BROWSE_URL=http://localhost:8080

Or via OpenClaw config (~/.openclaw/openclaw.json):

{
  "skills": {
    "entries": {
      "stealthy-auto-browse": {
        "env": {
          "STEALTHY_AUTO_BROWSE_URL": "http://localhost:8080"
        }
      }
    }
  }
}

3. Verify: curl $STEALTHY_AUTO_BROWSE_URL/health should return ok. VNC viewer at http://localhost:5900.

Key Concepts

Two input modes:

  1. System methods (system_click, system_type, mouse_move, send_key) — OS-level input via PyAutoGUI, completely undetectable. Use these for stealth.
  2. Playwright methods (click, fill, type) — CSS selector-based, faster but detectable by behavioral analysis.

Workflow for undetectable browsing:

  1. goto — navigate to page
  2. get_interactive_elements — find all clickable elements with x,y coordinates
  3. system_click — click at coordinates (real mouse movement)
  4. system_type — type text (real keystrokes)
  5. Take screenshot to verify — GET /screenshot/browser?whLargest=512

API Reference

All commands: POST $STEALTHY_AUTO_BROWSE_URL/ with JSON {"action": "<name>", ...params}.

Navigation

goto — Navigate to URL

{"action": "goto", "url": "https://example.com", "wait_until": "domcontentloaded"}

wait_until: domcontentloaded (default), load, networkidle

Undetectable Input (Prefer These)

system_click — Move mouse + click at coordinates

{"action": "system_click", "x": 500, "y": 300, "duration": 0.3}

mouse_move — Move mouse without clicking

{"action": "mouse_move", "x": 500, "y": 300, "duration": 0.5}

mouse_click — Click at position (current or specified)

{"action": "mouse_click", "x": 500, "y": 300}

system_type — Type text via OS keyboard

{"action": "system_type", "text": "hello world", "interval": 0.08}

send_key — Send keyboard key or combo

{"action": "send_key", "key": "enter"}
{"action": "send_key", "key": "ctrl+a"}

scroll — Scroll page (negative = down)

{"action": "scroll", "amount": -3}
{"action": "scroll", "amount": 5, "x": 500, "y": 300}

Playwright Input (Detectable)

click — Click by CSS selector

{"action": "click", "selector": "#submit-btn"}

fill — Fill input field (clears first)

{"action": "fill", "selector": "input[name='email']", "value": "user@example.com"}

type — Type into element with delay

{"action": "type", "selector": "#search", "text": "query", "delay": 0.05}

Page Inspection

get_interactive_elements — Get all clickable elements with coordinates

{"action": "get_interactive_elements", "visible_only": true}

Returns elements with tag, text, selector, x, y, w, h, visible. Use x and y with system_click for undetectable clicks.

get_text — Get visible text content

{"action": "get_text"}

get_html — Get full page HTML

{"action": "get_html"}

eval — Execute JavaScript

{"action": "eval", "expression": "document.title"}

Screenshots

GET /screenshot/browser — Browser viewport PNG GET /screenshot/desktop — Full desktop PNG

Resize query parameters (all optional):

| Parameter | Effect | |-----------|--------| | whLargest=512 | Set largest side to 512px, keep aspect ratio | | width=800 | Resize to 800px wide, keep aspect ratio | | height=300 | Resize to 300px tall, keep aspect ratio | | width=400&height=400 | Exact dimensions |

curl $STEALTHY_AUTO_BROWSE_URL/screenshot/browser?whLargest=512 -o screenshot.png

State & Utility

GET /state — Current URL, title, window offset GET /health — Returns "ok"

ping — Check connection

{"action": "ping"}

calibrate — Recalculate window offset for coordinate mapping

{"action": "calibrate"}

get_resolution — Get display resolution

{"action": "get_resolution"}

enter_fullscreen / exit_fullscreen — Toggle browser fullscreen

{"action": "enter_fullscreen"}
{"action": "exit_fullscreen"}

close — Shut down browser

{"action": "close"}

Container Options

# Custom resolution
docker run -d -p 8080:8080 -e XVFB_RESOLUTION=1280x720 psyb0t/stealthy-auto-browse

# Match timezone to IP location (important for stealth)
docker run -d -p 8080:8080 -e TZ=Europe/Bucharest psyb0t/stealthy-auto-browse

# Persistent browser profile (cookies, sessions, fingerprint)
docker run -d -p 8080:8080 -v ./profile:/userdata psyb0t/stealthy-auto-browse

# Open URL on startup
docker run -d -p 8080:8080 psyb0t/stealthy-auto-browse https://example.com

Example: Undetectable Login

API=$STEALTHY_AUTO_BROWSE_URL

# Navigate
curl -X POST $API -H "Content-Type: application/json" \
  -d '{"action": "goto", "url": "https://example.com/login"}'

# Find elements
curl -X POST $API -H "Content-Type: application/json" \
  -d '{"action": "get_interactive_elements"}'

# Click email field and type (undetectable)
curl -X POST $API -H "Content-Type: application/json" \
  -d '{"action": "system_click", "x": 400, "y": 200}'
curl -X POST $API -H "Content-Type: application/json" \
  -d '{"action": "system_type", "text": "user@example.com"}'

# Click password field and type
curl -X POST $API -H "Content-Type: application/json" \
  -d '{"action": "system_click", "x": 400, "y": 260}'
curl -X POST $API -H "Content-Type: application/json" \
  -d '{"action": "system_type", "text": "password123"}'

# Submit
curl -X POST $API -H "Content-Type: application/json" \
  -d '{"action": "system_click", "x": 400, "y": 320}'

# Verify with screenshot
curl "$API/screenshot/browser?whLargest=512" -o result.png

Response Format

{
  "success": true,
  "timestamp": 1234567890.123,
  "data": { ... },
  "error": "message if failed"
}

Tips

  1. Use get_interactive_elements to find coordinates before clicking
  2. Prefer system methods (system_click, system_type) for stealth
  3. Add delays between actions to appear human
  4. Use calibrate if coordinate clicks seem offset
  5. Match TZ to your IP's geographic location
  6. Use whLargest=512 on screenshots to keep them manageable
  7. Mount /userdata for persistent cookies, sessions, and fingerprint across restarts