Runs a real kasmweb/chrome Docker sidecar alongside OpenClaw. Principal logs in once via noVNC. Sessions saved permanently in a Docker volume. Agent accesses any authenticated platform from that point on — no credentials needed.

Three execution paths

Path	Use for	Tokens
OpenClaw native browser	Simple navigate / click / extract	Fewest
browser_control.py	AUDIT logging, workflows, CAPTCHA, Vision	Medium
noVNC (manual)	Initial login, 2FA, session renewal	None

Load only the path the task requires.

Key features

Persistent sessions — log in once, access forever
Autonomous CAPTCHA — CapSolver resolves reCAPTCHA / hCaptcha / Turnstile automatically
Residential proxy — Browserbase for sites that block VPS IPs
Claude Vision — analyze any page or image natively
Full logging — every action in AUDIT.md, errors in ERRORS.md, patterns in LEARNINGS.md
Self-improving — agent writes learnings immediately after every session

What gets downloaded at setup

What	Why
`kasmweb/chrome:1.15.0` (~2GB)	Real Chrome Desktop with noVNC — your authenticated sessions live here
`Playwright Chromium` (~100MB)	Headless fallback inside the OpenClaw container — used if Chrome CDP is unavailable
`requests + playwright` (Python)	Python dependencies for browser_control.py
`CapSolver extension` (optional)	Only if CAPSOLVER_API_KEY is set — autonomous CAPTCHA

Setup

Agent runs virtual_desktop.setup — downloads and installs everything, notifies the principal. Principal connects once via noVNC. Done.

Optional keys (add to .env)

CAPSOLVER_API_KEY=xxx      # Autonomous CAPTCHA (~$0.001/solve)
BROWSERBASE_API_KEY=xxx    # Residential proxy + stealth (free tier available)
ANTHROPIC_API_KEY=xxx      # Claude Vision (uses existing key if already set)

Files included

File	Role
`SKILL.md`	Full instructions for the agent
`browser_control.py`	Python helper — logging, CAPTCHA, Vision, workflows
`README.md`	This file
`CONFIGURATION.md`	noVNC access, .env variables, reset procedures