Openclaw Skill Browser Use

Autonomous browser automation for AI agents. Two tools: agent-browser (CLI Playwright for step-by-step control) and browser-use (Python autonomous agent that...

Security Scan

The skill appears to do what it claims (browser automation) but includes a system-level installer and runtime wrappers that auto-read OpenClaw config for API keys and write system-wide files — behaviors that are surprising and warrant careful review before install/use.

Purpose & Capability

The stated purpose (CLI Playwright control + autonomous Python agent) matches the provided files and commands. Requiring node/npm/python3/chromium/xvfb is coherent for browser automation. However the skill does not declare required environment variables in registry metadata even though the scripts expect OPENAI_API_KEY or ANTHROPIC_API_KEY (and will auto-detect them from /root/.openclaw/openclaw.json). That discrepancy (no declared env but runtime use of credentials) is a notable mismatch.

Instruction Scope

SKILL.md and the included scripts direct the agent to run arbitrary browser actions (open, click, eval JS) which is expected. The installer and wrappers (scripts/install.sh and scripts/browser-use-agent.sh) read /root/.openclaw/openclaw.json to extract API keys if env vars are not set — this is reading a central agent config file outside the skill's declared scope. The wrapper also prints/executes a temporary Python script using those keys. Accessing a root OpenClaw config for secrets is scope creep and potentially surprising.

Install Mechanism

There is no registry install spec, but an included scripts/install.sh performs apt-get installs, npm -g install agent-browser, pip installs of browser-use and langchain-* packages, and playwright browser installs. These are standard package sources (apt, npm, PyPI, Playwright) rather than arbitrary downloads, but the installer creates system-wide artifacts (/opt/browser-use, /usr/local/bin/browser-use-agent) and may require sudo. Global npm/pip installs and Playwright downloads should be reviewed before running.

Credentials

The skill will use LLM provider keys (OPENAI_API_KEY or ANTHROPIC_API_KEY) which are reasonable for an autonomous browsing agent, but the registry lists no required env vars. More importantly, both the wrapper and browser-use-agent.sh will attempt to pull keys from /root/.openclaw/openclaw.json if env vars are unset, which accesses central credentials. Reading that file can expose API keys stored by the platform — this is a high-sensitivity access that the skill did not declare.

Persistence & Privilege

Although always:false and model invocation is allowed (normal), the installer writes persistent, system-wide files: a venv in /opt/browser-use and a wrapper in /usr/local/bin/browser-use-agent. Combined with the auto-detection of keys from /root/.openclaw/openclaw.json, this persistent presence increases blast radius if keys are accessible. The installer also runs privileged package installs (apt-get) and global npm installs — running it as root or with sudo should be done cautiously.

What to consider before installing

What to consider before installing or running this skill: - The skill's functionality (headless browser automation + autonomous agent) is consistent with the files provided, but the scripts will try to read OpenClaw's central config (/root/.openclaw/openclaw.json) to auto-fill OPENAI_API_KEY or ANTHROPIC_API_KEY if you haven't exported them. That file can contain sensitive API keys for other services — decide whether you want this skill to access them. - The included installer performs system-wide changes: apt-get package installs, npm -g agent-browser, pip installs, Playwright browser downloads, creates /opt/browser-use and /usr/local/bin/browser-use-agent. Only run the installer on systems you control and understand; prefer running in a disposable VM/container or sandbox. - If you want to avoid the skill auto-reading platform credentials, set OPENAI_API_KEY or ANTHROPIC_API_KEY in the environment explicitly before running, or inspect and modify the wrapper/browser-use-agent.sh to remove the auto-detection lines that read /root/.openclaw/openclaw.json. - Review the npm package agent-browser and the PyPI package browser-use (and their dependencies) before global installation: npm packages can run postinstall scripts. If possible, install locally or pin versions. - Do not run the installer as root on a production host unless you accept the system-wide changes. Consider running the CLI-only agent-browser interactions without running the installer (if agent-browser is already present) or using a container. - If you want to proceed, audit scripts/install.sh and scripts/browser-use-agent.sh line-by-line (they are included) and confirm you accept creation of /opt and /usr/local artifacts and the reading of /root/.openclaw/openclaw.json.

Browser Use — Autonomous Browser Automation

Two complementary tools for browser automation:

Tool	Best for	How it works
agent-browser	Step-by-step control, scraping, form filling	CLI commands, you drive each action
browser-use	Complex autonomous tasks	Python agent that decides actions itself

Quick Start

agent-browser (recommended for most tasks)

# Navigate and inspect
agent-browser open "https://example.com"
agent-browser snapshot -i          # Get interactive elements with @refs

# Interact using refs
agent-browser click @e3            # Click element
agent-browser fill @e2 "text"      # Fill input (clears first)
agent-browser press Enter          # Press key

# Extract data
agent-browser get text @e1         # Get element text
agent-browser get attr @e1 href    # Get attribute
agent-browser screenshot /tmp/p.png # Screenshot

# Done
agent-browser close

browser-use (autonomous agent)

# Run a full autonomous browsing task
browser-use-agent "Find the pricing for Notion and compare plans"

The agent will navigate, click, read pages, and return a structured result.

agent-browser — Full Reference

Navigation

agent-browser open <url>           # Navigate to URL
agent-browser back                 # Go back
agent-browser forward              # Go forward
agent-browser reload               # Reload page
agent-browser close                # Close browser

Snapshot (page analysis)

agent-browser snapshot             # Full accessibility tree
agent-browser snapshot -i          # Interactive elements only (recommended)
agent-browser snapshot -c          # Compact output
agent-browser snapshot -d 3        # Limit depth to 3
agent-browser snapshot -s "#main"  # Scope to CSS selector
agent-browser snapshot -i --json   # JSON output for parsing

Interactions (use @refs from snapshot)

agent-browser click @e1            # Click
agent-browser dblclick @e1         # Double-click
agent-browser fill @e2 "text"      # Clear and type (use this for inputs)
agent-browser type @e2 "text"      # Type without clearing
agent-browser press Enter          # Press key
agent-browser press Control+a      # Key combination
agent-browser hover @e1            # Hover
agent-browser check @e1            # Check checkbox
agent-browser uncheck @e1          # Uncheck checkbox
agent-browser select @e1 "value"   # Select dropdown option
agent-browser scroll down 500      # Scroll page
agent-browser scrollintoview @e1   # Scroll element into view
agent-browser drag @e1 @e2         # Drag and drop
agent-browser upload @e1 file.pdf  # Upload files

Extract Data

agent-browser get text @e1         # Get element text
agent-browser get html @e1         # Get innerHTML
agent-browser get value @e1        # Get input value
agent-browser get attr @e1 href    # Get attribute
agent-browser get title            # Page title
agent-browser get url              # Current URL
agent-browser get count ".item"    # Count matching elements

Wait

agent-browser wait @e1             # Wait for element
agent-browser wait 2000            # Wait milliseconds
agent-browser wait --text "Done"   # Wait for text to appear
agent-browser wait --url "/dash"   # Wait for URL pattern
agent-browser wait --load networkidle  # Wait for network idle

Screenshots, PDF & Recording

agent-browser screenshot path.png      # Save screenshot
agent-browser screenshot --full        # Full page screenshot
agent-browser pdf output.pdf           # Save as PDF
agent-browser record start ./demo.webm # Start recording
agent-browser record stop              # Stop and save

Sessions (parallel browsers)

agent-browser --session s1 open "https://site1.com"
agent-browser --session s2 open "https://site2.com"
agent-browser session list

State (persist auth/cookies)

agent-browser state save auth.json     # Save session (cookies, storage)
agent-browser state load auth.json     # Restore session

Cookies & Storage

agent-browser cookies                  # Get all cookies
agent-browser cookies set name value   # Set cookie
agent-browser cookies clear            # Clear cookies
agent-browser storage local            # Get all localStorage
agent-browser storage local set k v    # Set value

Tabs & Frames

agent-browser tab                      # List tabs
agent-browser tab new [url]            # New tab
agent-browser tab 2                    # Switch to tab
agent-browser frame "#iframe"          # Switch to iframe
agent-browser frame main               # Back to main frame

Browser Settings

agent-browser set viewport 1920 1080
agent-browser set device "iPhone 14"
agent-browser set geo 37.7749 -122.4194
agent-browser set offline on
agent-browser set media dark

JavaScript

agent-browser eval "document.title"    # Run JS in page context

browser-use — Autonomous Agent

For complex tasks where you want the agent to figure out the browsing steps:

browser-use-agent "Your task description here"

Custom Script (advanced)

# Run via: /opt/browser-use/bin/python3 script.py
import asyncio, os
from browser_use import Agent, Browser
from langchain_anthropic import ChatAnthropic

async def run():
    browser = Browser()
    llm = ChatAnthropic(
        model='claude-sonnet-4-20250514',
        api_key=os.environ['ANTHROPIC_API_KEY']
    )
    agent = Agent(
        task="Compare pricing on 3 competitor sites",
        llm=llm,
        browser=browser,
    )
    result = await agent.run(max_steps=15)
    await browser.close()
    return result

asyncio.run(run())

You can swap the LLM for any langchain-compatible model (OpenAI, Anthropic, etc).

Standard Workflow

# 1. Open page
agent-browser open "https://example.com"

# 2. Snapshot to see what's on the page
agent-browser snapshot -i

# 3. Interact with elements using @refs from snapshot
agent-browser fill @e1 "search query"
agent-browser click @e2

# 4. Wait for new page to load
agent-browser wait --load networkidle

# 5. Re-snapshot (refs change after navigation!)
agent-browser snapshot -i

# 6. Extract what you need
agent-browser get text @e5

# 7. Close when done
agent-browser close

Important Rules

1. Always snapshot -i after navigation — refs change on every page load
2. Use fill not type for inputs — fill clears existing text first
3. Wait after clicks that trigger navigation — wait --load networkidle
4. Close the browser when done — agent-browser close
5. Google/Bing block headless browsers (CAPTCHA) — use DuckDuckGo or web_search instead
6. Save auth state for sites requiring login — state save/load
7. Use --json when you need machine-parseable output
8. Use sessions for parallel browsing — --session <name>

Troubleshooting

• Element not found: Re-run snapshot -i to get current refs
• Page not loaded: Add wait --load networkidle after navigation
• CAPTCHA on search engines: Use DuckDuckGo or the web_search tool instead
• Auth expired: Re-login and state save again
• Display errors: The install script sets up Xvfb for headless rendering