ClawBrowser

A stealth browser with a localhost HTTP control API, built for AI agents.

No Selenium. No Playwright. No Chrome DevTools Protocol. No automation fingerprints.

Just PyQt6 WebEngine (real Chromium) + a simple REST API on localhost.

Why

Standard automation tools (Selenium, Playwright, CDP) leave detectable signatures. Sites like REA, LinkedIn, and anything running Kasada/PerimeterX will block them. ClawBrowser presents as a normal user-launched Chromium process — because it is one.

Built for AI agents that need to browse the web without getting blocked. Also usable as a fast, no-telemetry browser for humans.

Install

Requirements

• Python 3.10–3.12 (PyQt6-WebEngine is not yet available for 3.13+)
• Linux (tested on Ubuntu 22.04 / Pop!_OS). macOS should work. Windows untested.

Check your Python version

python3 --version

If you're on 3.13+ (e.g. via Homebrew), find your system Python:

ls /usr/bin/python3*
# Use /usr/bin/python3 explicitly below

Install dependencies

pip install PyQt6 PyQt6-WebEngine
# or explicitly:
/usr/bin/python3 -m pip install PyQt6 PyQt6-WebEngine

Clone

git clone https://github.com/anythingwithawire/clawbrowser
cd clawbrowser
chmod +x claw.sh

Run

python3 browser.py              # default port 8765
python3 browser.py --port 9000  # custom port
CLAW_PORT=9000 python3 browser.py

# Linux with multiple Python versions — use system Python explicitly:
/usr/bin/python3 browser.py

On Linux, a DISPLAY must be set (for headless/server use, set DISPLAY=:1 if you have a virtual framebuffer):

DISPLAY=:1 /usr/bin/python3 browser.py &
sleep 4  # wait for Qt + API to initialise

Check it's up:

curl -sf http://127.0.0.1:8765/status

Control via CLI

./claw.sh status                   # browser status + ads blocked
./claw.sh go https://example.com   # navigate
./claw.sh shot [file]              # screenshot (default: /tmp/claw_shot.png)
./claw.sh dom [chars]              # page HTML
./claw.sh text [chars]             # visible text only
./claw.sh click "button.submit"    # click by CSS selector
./claw.sh click 450,300            # click by coordinates
./claw.sh type "#search" "hello"   # type into input
./claw.sh js "document.title"      # run JavaScript
./claw.sh links                    # list all links
./claw.sh forms                    # list all form elements
./claw.sh tabs                     # list open tabs
./claw.sh tab new https://...      # open new tab
./claw.sh wait [secs]              # wait then show status

HTTP API

All endpoints on http://127.0.0.1:8765 (or your configured port):

Method	Endpoint	Body	Description
GET	/status	—	Browser state
POST	/navigate	{"url": "..."}	Navigate
GET	/screenshot	—	PNG bytes
GET	/dom	—	{"html": "..."}
POST	/eval	{"script": "..."}	Run JS, returns {"result": ...}
POST	/click	{"selector": "..."} or {"x": n, "y": n}	Click
POST	/type	{"selector": "...", "text": "..."}	Type
GET	/links	—	All <a> elements
GET	/forms	—	All form inputs
GET	/tabs	—	Tab list
POST	/tab/new	{"url": "..."}	New tab
POST	/tab/close	{"index": n}	Close tab
POST	/tab/switch	{"index": n}	Switch tab
GET	/bookmarks	—	Bookmark list
POST	/bookmarks	{"url": "...", "title": "..."}	Add bookmark
DELETE	/bookmarks	{"url": "..."}	Remove bookmark
GET	/history	?q=&limit=	Search history

Ad Blocking

Domain blocklist at ~/.clawbrowser/blocked_domains.txt — one domain per line.

OpenClaw Integration

If you're using OpenClaw, drop SKILL.md into your workspace skills directory:

mkdir -p ~/.openclaw/workspace/skills/clawbrowser
cp SKILL.md ~/.openclaw/workspace/skills/clawbrowser/

OpenClaw will auto-discover the skill and use ClawBrowser when bot detection is a concern.

Features

• Real Chromium rendering via Qt WebEngine
• Tabbed UI (usable by humans too)
• Ad blocking via domain blocklist
• Bookmarks + history (persisted to ~/.clawbrowser/)
• Logins persist across restarts (cookies in ~/.clawbrowser/storage/)
• No telemetry, no Google sync, no crash reporting
• No webdriver flags — invisible to bot detection
• Localhost-only API (not accessible from network)

Notes

• If browser.py is already running, claw.sh will connect to it. No need to restart.
• Kill with: pkill -f "browser.py"
• GBM/Vulkan warnings on startup are harmless — ignore them.
• Screenshot → vision model is the most reliable way to understand page state when selectors fail.