╔═══════════════════════════════════════════════════════════════════╗
║                                                                   ║
║   ██╗    ██╗███████╗██████╗  ██████╗██╗     ██╗                   ║
║   ██║    ██║██╔════╝██╔══██╗██╔════╝██║     ██║                   ║
║   ██║ █╗ ██║█████╗  ██████╔╝██║     ██║     ██║                   ║
║   ██║███╗██║██╔══╝  ██╔══██╗██║     ██║     ██║                   ║
║   ╚███╔███╔╝███████╗██████╔╝╚██████╗███████╗██║                   ║
║    ╚══╝╚══╝ ╚══════╝╚═════╝  ╚═════╝╚══════╝╚═╝                   ║
║                                                                   ║
║   Browse the web like it's 1994 — for AI agents                   ║
║                                                                   ║
╚═══════════════════════════════════════════════════════════════════╝

WebCLI is a text-mode web browser designed for AI agents.

It converts modern websites into BBS-style, numbered text menus, so LLMs can navigate the web using the same format they already reason in: text.

No screenshots.
No raw HTML dumps.
No brittle selectors.

Working with browsers + LLMs today is painful.

Most agent setups:

• rely on screenshots + vision models
• or feed raw HTML full of JavaScript, hidden nodes, and noise

Both approaches are:

• expensive
• slow
• hard for models to reason about
• fragile across sites

I built WebCLI after repeatedly hitting these problems while experimenting with browser-based agents.

This project explores a different idea: what if the browser spoke the same language as the agent?

Turns this:

<a href="/story/123" class="storylink">How I Built X</a>
<span class="score">172 points</span>
<!-- ... 10,000 more lines ... -->

Into this:

[L10] How I Built X
172 points by developer | 3 hours ago
[L15] 113 comments

This keeps interaction deterministic, readable, and cheap.

git clone https://github.com/andycufari/webcli.git
cd webcli
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
playwright install chromium

Add to your config:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "webcli": {
      "command": "/path/to/webcli/venv/bin/python",
      "args": ["/path/to/webcli/mcp_server.py"]
    }
  }
}

Restart Claude. Then:

"Go to news.ycombinator.com and tell me the top stories"

python webcli.py

🌐 > goto news.ycombinator.com
🌐 > click L12
🌐 > fill I1 "search term"
🌐 > back
🌐 > quit

from webcli import CLIBrowser
import asyncio

async def main():
    browser = CLIBrowser(headless=True)
    await browser.start()
    await browser.goto("https://amazon.com")
    await browser.fill("I1", "mechanical keyboard")
    await browser.click("B1")
    print(browser.render())
    await browser.close()

asyncio.run(main())

[L1], [L2]...  = Links
[B1], [B2]...  = Buttons  
[I1], [I2]...  = Inputs
[S1], [S2]...  = Selects

LLM ──► MCP Server ──► CLIBrowser ──► browser-use ──► Playwright ──► Chromium

• Playwright runs headless Chrome
• browser-use extracts interactive elements from DOM
• CLIBrowser renders them as numbered text menus
• MCP Server exposes tools to Claude/other LLMs

• Search engines (Google, Bing) block automated browsers
• Can't solve CAPTCHAs
• Some SPAs need scrolling to load content
• No file uploads yet

• E-commerce: search products, compare prices
• Research: navigate docs, extract info
• Forms: fill repetitive web forms
• Testing: web tests without brittle selectors

• browser-use
• Playwright
• MCP

MIT © Andy Cufari / CM64.studio

Built in Buenos Aires 🇦🇷