browser-use（浏览器使用）

Browser Automation with browser-use

All commands use the self-contained wrapper bu.sh, which:

• Auto-bootstraps the local .venv on first run (downloads deps via uv — takes ~60s once)
• Sets PYTHONUTF8=1 automatically (handles emoji/Unicode on Windows)
• Keeps a persistent browser daemon for ~50ms latency per call

Shorthand — in all examples, bu means:

bash ~/.claude/skills/browser-use/bu.sh

Prerequisites

• uv must be installed (uv --version)
• First run will install browser-use from GitHub into ~/.claude/skills/browser-use/.venv

Core Workflow

1. Navigate: bu open <url> — starts browser + daemon if needed
2. Inspect: bu state — returns clickable elements with indices
3. Interact: use indices from state (bu click 5, bu input 3 "text")
4. Verify: bu state or bu screenshot to confirm
5. Repeat: browser stays open between commands
6. Cleanup: bu close when done

Browser Modes

bu open <url>                         # Default: headless Chromium
bu --headed open <url>                # Visible window
bu --profile "Default" open <url>     # Real Chrome with Default profile (existing logins/cookies)
bu --profile "Profile 1" open <url>   # Real Chrome with named profile
bu --connect open <url>               # Auto-discover running Chrome via CDP
bu --cdp-url ws://localhost:9222/... open <url>  # Connect via specific CDP URL

--connect, --cdp-url, and --profile are mutually exclusive.

Commands

# Navigation
bu open <url>                    # Navigate to URL
bu back                          # Go back in history
bu scroll down                   # Scroll down (--amount N for pixels)
bu scroll up                     # Scroll up
bu switch <tab>                  # Switch to tab by index
bu close-tab [tab]               # Close tab (current if no index)

# Page State — always run state first to get element indices
bu state                         # URL, title, clickable elements with indices
bu screenshot [path.png]         # Screenshot (base64 if no path, --full for full page)

# Interactions — use indices from state
bu click <index>                 # Click element by index
bu click <x> <y>                 # Click at pixel coordinates
bu type "text"                   # Type into focused element
bu input <index> "text"          # Click element, then type
bu keys "Enter"                  # Send keyboard keys (also "Control+a", etc.)
bu select <index> "option"       # Select dropdown option
bu upload <index> <path>         # Upload file to file input
bu hover <index>                 # Hover over element
bu dblclick <index>              # Double-click element
bu rightclick <index>            # Right-click element

# Data Extraction
bu eval "js code"                # Execute JavaScript, return result
bu get title                     # Page title
bu get html [--selector "h1"]    # Page HTML (or scoped to selector)
bu get text <index>              # Element text content
bu get value <index>             # Input/textarea value
bu get attributes <index>        # Element attributes
bu get bbox <index>              # Bounding box (x, y, width, height)

# Wait
bu wait selector "css"           # Wait for element (--state visible|hidden|attached|detached, --timeout ms)
bu wait text "text"              # Wait for text to appear

# Cookies
bu cookies get [--url <url>]     # Get cookies (optionally filtered)
bu cookies set <name> <value>    # Set cookie (--domain, --secure, --http-only, --same-site, --expires)
bu cookies clear [--url <url>]   # Clear cookies
bu cookies export <file>         # Export to JSON
bu cookies import <file>         # Import from JSON

# Python — persistent session with browser access
bu python "code"                 # Execute Python (variables persist across calls)
bu python --file script.py       # Run file
bu python --vars                 # Show defined variables
bu python --reset                # Clear namespace

# Session
bu close                         # Close browser and stop daemon
bu sessions                      # List active sessions
bu close --all                   # Close all sessions

# Recipes — store and replay DOM access patterns (no screenshot on repeat runs)
bu recipe list                   # Show all installed recipes + run history
bu recipe run <name>             # Execute a recipe directly against the DOM
bu recipe run <name> --connect   # Override auth: use CDP connect mode
bu recipe run <name> --headed    # Override: show browser window (debug)
bu recipe show <name>            # Print recipe JSON
bu recipe save <name> --file <path>  # Install a recipe JSON file
bu recipe delete <name>          # Remove a recipe

The Python browser object provides: browser.url, browser.title, browser.html, browser.goto(url), browser.back(), browser.click(index), browser.type(text), browser.input(index, text), browser.keys(keys), browser.upload(index, path), browser.screenshot(path), browser.scroll(direction, amount), browser.wait(seconds).

Cloud API

bu cloud connect                 # Provision cloud browser and connect
bu cloud connect --timeout 120 --proxy-country US  # With options
bu cloud login <api-key>         # Save API key (or set BROWSER_USE_API_KEY)
bu cloud logout                  # Remove API key
bu cloud v2 GET /browsers        # REST passthrough (v2 or v3)
bu cloud v2 POST /tasks '{"task":"...","url":"..."}'
bu cloud v2 poll <task-id>       # Poll task until done

Common Workflows

Authenticated Browsing (existing Chrome session)

bu --connect open https://gmail.com   # Reuse logged-in Chrome via CDP
bu --connect state
bu --connect eval "document.title"

Requires Chrome launched with --remote-debugging-port=9222.

Authenticated Browsing (Chrome profile)

bu profile list                                # Check available profiles
bu --profile "Default" open https://github.com # Already logged in

Extracting Data via JavaScript

bu --connect open https://x.com/notifications
bu --connect eval "Array.from(document.querySelectorAll('[data-testid=\"notification\"]')).slice(0,10).map(n => n.innerText).join('\n---\n')"

Exposing Local Dev Servers

bu tunnel 3000                             # → https://abc.trycloudflare.com
bu open https://abc.trycloudflare.com

Command Chaining

bu open https://example.com && bu state
bu input 5 "user@example.com" && bu input 6 "password" && bu click 7

Chain when you don't need intermediate output. Run separately when you need to parse state to discover indices first.

Global Options

Tips

1. Always run state first to see available elements and their indices
2. Use --headed for debugging to see what the browser is doing
3. Sessions persist — browser stays open between commands
4. eval with JS is the most powerful extraction method for complex pages

Troubleshooting

• First run slow? Normal — downloading ~200MB of deps into .venv. Subsequent runs are instant.
• Browser won't start? bu close then bu --headed open <url>
• Element not found? bu scroll down then bu state
• Unicode errors? Already handled — bu.sh sets PYTHONUTF8=1 automatically

Cleanup

bu close           # Close browser session
bu tunnel stop --all  # Stop tunnels (if any)

Recipe System

Recipes store DOM access patterns so repeated tasks skip visual discovery entirely.

First run: discover selectors manually → write a recipe JSON Repeat runs: bu recipe run <name> executes JS directly — no screenshot, no element scanning

Recipe JSON format

{
  "name": "my_recipe",
  "version": 1,
  "description": "What this recipe does",
  "url": "https://example.com/page",

  "auth": {
    "mode": "connect",      // "connect" | "profile" | "headless"
    "profile": null,        // Chrome profile name (mode=profile only)
    "headed": false,
    "session": "default"
  },

  "steps": [
    {
      "id": "nav",
      "type": "navigate",
      "url": "https://example.com/page",
      "wait_for": ".main-content",
      "wait_timeout_ms": 10000
    },
    {
      "id": "extract",
      "type": "eval",
      "js": "document.querySelector('.result')?.innerText",
      "output_var": "result",
      "fallback_selectors": [".alt-result", "[data-result]"]
    }
  ],

  "output": { "format": "text", "var": "result" },
  "metadata": { "created": "2026-01-01", "last_run": null, "run_count": 0, "broken_step": null }
}

Step types

Self-healing fallback

When an eval step fails:

1. Each fallback_selectors entry is tried as document.querySelector('SEL')?.innerText
2. First match wins — logged to stderr, recipe continues
3. If all fallbacks fail: screenshot saved to recipes/.broken_<name>_<step>.png, broken_step set in metadata, exit 1

Sample recipe

See recipes/x_notifications.json — fetches X.com notifications using CDP connection.

Platform support

Works on Linux, macOS, and Windows (Git Bash or WSL required on Windows). bash must be available in PATH — it is on all standard configurations.