browse

Browser automation CLI with autonomous agent, multi-tab, cookies, and CAPTCHA solving. One Python script, connects to existing Chrome via CDP or launches headless browser.

$ browse task "Go to google.com and search for weather in Berlin"

mode: cdp
task: Go to google.com and search for weather in Berlin
max_steps: 30

[step 1] thinking... input 5 weather in Berlin
[step 2] thinking... click 9
[step 3] thinking... done The current temperature in Berlin is 6 C.

Task complete: The current temperature in Berlin is 6 C.

Install

curl -fsSL https://raw.githubusercontent.com/buildingopen/browse/main/browse -o ~/.local/bin/browse
chmod +x ~/.local/bin/browse

Requires: Python 3.8+, playwright (pip install playwright)

Commands

Command	Description
browse open <url>	Navigate to URL
browse state	Show numbered clickable elements
browse click <n>	Click element by index
browse input <n> <text>	Clear field and type
browse select <n> <value>	Select dropdown option
browse type <text>	Type into focused element
browse screenshot [file]	Take screenshot
browse extract	Extract page text
browse tabs	List all open tabs
browse tab <n>	Switch to tab
browse newtab <url>	Open URL in new tab
browse closetab [n]	Close tab
browse cookies save [file]	Save cookies to file
browse cookies load [file]	Restore cookies
browse solve	Solve CAPTCHA via 2captcha
browse task "<goal>"	Autonomous agent mode
browse close	Close browser

Autonomous Agent

Give it a goal, it figures out what to click:

browse task "Search Google for flights from NYC to London"
browse task "Fill out the signup form with name John and email john@test.com"
browse task "Go to GitHub trending and list the top 3 repos" --max-steps 10

The agent uses vision + DOM: sends both a screenshot and the element list to Gemini, so it understands visual layout, images, and dynamic content. Includes error recovery (retries on failures, stops after 3 consecutive errors).

Multi-Tab

browse newtab https://example.com    # Open in new tab
browse tabs                          # List tabs with indices
browse tab 2                         # Switch to tab
browse closetab 0                    # Close tab

Cookie Persistence

browse cookies save                  # Save to default file
browse cookies save ~/my-cookies.json
browse cookies load                  # Restore cookies

CAPTCHA Solving

export TWOCAPTCHA_KEY=your_key
browse solve                         # Solve reCAPTCHA/hCaptcha on current page

Detects reCAPTCHA and hCaptcha via data-sitekey, submits to 2captcha, polls for solution, injects the token.

Connection

Tries CDP first (BROWSE_CDP_URL, default http://localhost:9222), falls back to headless browser. On a server with Chrome running, it connects automatically.

Environment Variables

Variable	Default	Description
BROWSE_CDP_URL	http://localhost:9222	CDP endpoint
GEMINI_API_KEY	-	For autonomous agent
GOOGLE_API_KEY	-	Alternative to GEMINI_API_KEY
TWOCAPTCHA_KEY	-	For CAPTCHA solving

Why

browser-use (82k stars) has the same concept but its BrowserSession watchdog layer fails on headless Linux. browse uses Playwright directly.

License

MIT