agent-browser（Agent 浏览器自动化）

Agent Browser Skill

Fast browser automation CLI for AI agents. Use this to verify web UI changes, test interactions, and capture screenshots.

When to Use

• Verify visual changes after modifying frontend code
• Test form interactions and user flows
• Capture screenshots for documentation or debugging
• Inspect rendered HTML/accessibility tree
• Debug why something isn't working in the browser

Core Workflow

1. Open a URL

agent-browser open http://localhost:4321/

2. Take a Snapshot

The snapshot command returns an accessibility tree with refs (@e1, @e2, etc.) that you can use for interactions:

agent-browser snapshot -i    # -i = interactive elements only (recommended)
agent-browser snapshot -c    # -c = compact (removes empty structural elements)
agent-browser snapshot -i -c # Both flags work together

3. Interact Using Refs

Use the @ref values from the snapshot to interact with elements:

agent-browser click @e5           # Click element
agent-browser fill @e3 "hello"    # Clear and type
agent-browser type @e3 "world"    # Type without clearing
agent-browser select @e7 "option" # Select dropdown
agent-browser check @e9           # Check checkbox

4. Screenshot

agent-browser screenshot              # Viewport only
agent-browser screenshot --full       # Full page
agent-browser screenshot output.png   # Save to file

ALWAYS check the image size before attempting to load. If it is larger than 2MB, process it using a tool such as sips or ImageMagick to reduce the size. Large files cannot be loaded by your tools.

Common Patterns

Form Verification

agent-browser open http://localhost:4321/_emdash/api/setup/dev-bypass?redirect=/_emdash/admin/content/posts/new
agent-browser snapshot -i   # Check what fields are visible
agent-browser fill @e3 "Test Title"
agent-browser fill @e5 "Test content here"
agent-browser click @e7     # Save button

Get Element Info

agent-browser get text @e1      # Get text content
agent-browser get html @e1      # Get inner HTML
agent-browser get value @e2     # Get input value
agent-browser get url           # Current URL
agent-browser get title         # Page title
agent-browser get count "button" # Count matching elements

Check Element State

agent-browser is visible @e1
agent-browser is enabled @e2
agent-browser is checked @e3

Find by Role/Label

agent-browser find role button click --name "Submit"
agent-browser find label "Email" fill "test@example.com"
agent-browser find placeholder "Search..." type "query"

Sessions

Sessions keep browser state (cookies, storage) between commands:

# Use named session (persists until closed)
agent-browser --session mytest open http://localhost:4321
agent-browser --session mytest snapshot -i

# Or set via environment
export AGENT_BROWSER_SESSION=mytest
agent-browser open http://localhost:3000

Useful Options

Debugging

agent-browser --headed open http://localhost:3000  # See what's happening
agent-browser console                               # View console logs
agent-browser errors                                # View page errors
agent-browser highlight @e5                         # Highlight element
agent-browser eval "document.title"                 # Run JS

Tips

1. Always snapshot first - Get refs before interacting
2. Use -i flag - Interactive-only snapshots are much cleaner
3. Wait when needed - Use wait <ms> or wait <selector> after actions that trigger loading
4. Sessions for auth - Use named sessions to persist login state
5. Headed for debugging - Use --headed when things aren't working as expected