Browser Agent — Claude Code Skill

Give Claude Code full browser control. Navigate pages, click buttons, fill forms, take screenshots, and read page content — all through a real Chromium browser.

What It Does

• Launches a real Chromium browser controlled via Chrome DevTools Protocol (CDP)
• Claude can navigate, click, type, screenshot, and read any web page
• Persistent login profile — log in once, stay logged in across sessions
• Optional anti-detection mode (Patchright) for sites with bot protection
• Accessibility tree snapshots for token-efficient page understanding

Requirements

• Python 3.8+
• macOS, Linux, or Windows

Installation

1. Copy the tool

mkdir -p ~/.browser-tool
cp browser_tool.py ~/.browser-tool/browser_tool.py

2. Install dependencies

python3 ~/.browser-tool/browser_tool.py install

This installs Playwright and downloads Chromium automatically.

Optional — anti-detection mode (for Cloudflare/DataDome protected sites):

python3 ~/.browser-tool/browser_tool.py install --with-patchright

3. Add the skill to Claude Code

Copy the skill file to your Claude Code commands directory:

mkdir -p ~/.claude/commands
cp browser.md ~/.claude/commands/browser.md

4. Done

Open Claude Code and type /browser to activate browser control. Or just ask Claude to "open a browser and go to example.com" — it will know what to do.

How It Works

You (Claude Code) ──> browser_tool.py ──> CDP ──> Chromium
                                                    ↑
                                          Stays open as a
                                          detached process

1. launch starts Chromium with --remote-debugging-port=9222 as a detached process
2. Every other command (navigate, click, screenshot...) connects via CDP, does its work, disconnects
3. The browser stays open independently — no process blocking
4. Login sessions persist in ~/.browser-tool/profile/

Quick Reference

Command	What it does
launch	Start Chromium (or --patchright for anti-detection)
navigate <url>	Go to a URL
click <selector>	Click an element
type <selector> <text>	Type into a field (--clear to replace)
snapshot	Get accessibility tree (best for AI reasoning)
screenshot	Save viewport screenshot
html	Get page HTML
evaluate <js>	Run JavaScript
wait --selector <sel>	Wait for element to appear
pages	List open tabs
status	Check if browser is running
close	Kill the browser

Tips

• Use snapshot over screenshot when Claude needs to understand page structure — it's faster and more token-efficient
• Use --patchright when a site has Cloudflare, DataDome, or similar bot detection
• Logins persist — once you log into a site, you stay logged in across Claude Code sessions
• Screenshots are saved to /tmp/browser_screenshots/ and auto-opened in Preview (macOS)