agent-browser

name	agent-browser
slug	agent-browser
version	1.1.0
description	Self-contained browser automation. Navigates websites, fills forms, takes screenshots, extracts data. Runtime auto-installs on first use — no npm, Node.js, or Chrome required on the host.
setup	bootstrap.sh
allowed-tools	Bash(agent-browser:*)

Browser Automation with agent-browser

Runtime Setup (MANDATORY — run before any command)

Before executing ANY agent-browser command, you MUST ensure the runtime is installed:

1. Check if runtime/version.txt exists in this skill's directory
2. If it does NOT exist, run the bootstrap script:
   • macOS / Linux: bash "<SKILL_DIR>/bootstrap.sh"
   • Windows: powershell -ExecutionPolicy Bypass -File "<SKILL_DIR>/bootstrap.ps1" This downloads ~230MB runtime (one-time). Tell the user: "正在首次准备浏览器环境，请稍候..."
3. After bootstrap, ALL commands must go through the wrapper:
   • macOS / Linux: "<SKILL_DIR>/agent-browser" <command> [args...]
   • Windows: "<SKILL_DIR>/agent-browser.cmd" <command> [args...]

Replace <SKILL_DIR> with the absolute path to this skill's directory.

NEVER call agent-browser directly without the skill directory prefix. Always use "<SKILL_DIR>/agent-browser" (macOS/Linux) or "<SKILL_DIR>/agent-browser.cmd" (Windows).

Quick start

"<SKILL_DIR>/agent-browser" open <url>        # Navigate to page
"<SKILL_DIR>/agent-browser" snapshot -i       # Get interactive elements with refs
"<SKILL_DIR>/agent-browser" click @e1         # Click element by ref
"<SKILL_DIR>/agent-browser" fill @e2 "text"   # Fill input by ref
"<SKILL_DIR>/agent-browser" close             # Close browser

Core workflow

1. Navigate: "<SKILL_DIR>/agent-browser" open <url>
2. Snapshot: "<SKILL_DIR>/agent-browser" snapshot -i (returns elements with refs like @e1, @e2)
3. Interact using refs from the snapshot
4. Re-snapshot after navigation or significant DOM changes

Command Reference

Below, AB is shorthand for "<SKILL_DIR>/agent-browser". Replace <SKILL_DIR> with the absolute path to this skill's directory in every invocation.

Navigation

"<SKILL_DIR>/agent-browser" open <url>
"<SKILL_DIR>/agent-browser" back
"<SKILL_DIR>/agent-browser" forward
"<SKILL_DIR>/agent-browser" reload
"<SKILL_DIR>/agent-browser" close
"<SKILL_DIR>/agent-browser" connect 9222    # Connect via CDP port

Snapshot (page analysis)

"<SKILL_DIR>/agent-browser" snapshot            # Full accessibility tree
"<SKILL_DIR>/agent-browser" snapshot -i         # Interactive elements only (recommended)
"<SKILL_DIR>/agent-browser" snapshot -c         # Compact output
"<SKILL_DIR>/agent-browser" snapshot -d 3       # Limit depth to 3
"<SKILL_DIR>/agent-browser" snapshot -s "#main" # Scope to CSS selector

Interactions (use @refs from snapshot)

"<SKILL_DIR>/agent-browser" click @e1
"<SKILL_DIR>/agent-browser" dblclick @e1
"<SKILL_DIR>/agent-browser" fill @e2 "text"     # Clear and type
"<SKILL_DIR>/agent-browser" type @e2 "text"     # Type without clearing
"<SKILL_DIR>/agent-browser" press Enter
"<SKILL_DIR>/agent-browser" hover @e1
"<SKILL_DIR>/agent-browser" check @e1
"<SKILL_DIR>/agent-browser" uncheck @e1
"<SKILL_DIR>/agent-browser" select @e1 "value"
"<SKILL_DIR>/agent-browser" scroll down 500
"<SKILL_DIR>/agent-browser" scrollintoview @e1
"<SKILL_DIR>/agent-browser" drag @e1 @e2
"<SKILL_DIR>/agent-browser" upload @e1 file.pdf

Get information

"<SKILL_DIR>/agent-browser" get text @e1
"<SKILL_DIR>/agent-browser" get html @e1
"<SKILL_DIR>/agent-browser" get value @e1
"<SKILL_DIR>/agent-browser" get attr @e1 href
"<SKILL_DIR>/agent-browser" get title
"<SKILL_DIR>/agent-browser" get url
"<SKILL_DIR>/agent-browser" get count ".item"

Screenshots & PDF

"<SKILL_DIR>/agent-browser" screenshot
"<SKILL_DIR>/agent-browser" screenshot path.png
"<SKILL_DIR>/agent-browser" screenshot --full    # Full page
"<SKILL_DIR>/agent-browser" pdf output.pdf

Wait

"<SKILL_DIR>/agent-browser" wait @e1
"<SKILL_DIR>/agent-browser" wait 2000
"<SKILL_DIR>/agent-browser" wait --text "Success"
"<SKILL_DIR>/agent-browser" wait --url "**/dashboard"
"<SKILL_DIR>/agent-browser" wait --load networkidle

Semantic locators

"<SKILL_DIR>/agent-browser" find role button click --name "Submit"
"<SKILL_DIR>/agent-browser" find text "Sign In" click
"<SKILL_DIR>/agent-browser" find label "Email" fill "[e-mail]"

Cookies & Storage

"<SKILL_DIR>/agent-browser" cookies
"<SKILL_DIR>/agent-browser" cookies set name value
"<SKILL_DIR>/agent-browser" cookies clear
"<SKILL_DIR>/agent-browser" storage local

Tabs

"<SKILL_DIR>/agent-browser" tab
"<SKILL_DIR>/agent-browser" tab new [url]
"<SKILL_DIR>/agent-browser" tab 2
"<SKILL_DIR>/agent-browser" tab close

Dialogs

"<SKILL_DIR>/agent-browser" dialog accept [text]
"<SKILL_DIR>/agent-browser" dialog dismiss

JavaScript

"<SKILL_DIR>/agent-browser" eval "document.title"

Global options

"<SKILL_DIR>/agent-browser" --session <name> ...    # Isolated browser session
"<SKILL_DIR>/agent-browser" --json ...              # JSON output for parsing
"<SKILL_DIR>/agent-browser" --headed ...            # Show browser window
"<SKILL_DIR>/agent-browser" --proxy <url> ...       # Use proxy server

Example: Form submission

"<SKILL_DIR>/agent-browser" open https://example.com/form
"<SKILL_DIR>/agent-browser" snapshot -i
"<SKILL_DIR>/agent-browser" fill @e1 "[e-mail]"
"<SKILL_DIR>/agent-browser" fill @e2 "password123"
"<SKILL_DIR>/agent-browser" click @e3
"<SKILL_DIR>/agent-browser" wait --load networkidle
"<SKILL_DIR>/agent-browser" snapshot -i

Example: Authentication with saved state

"<SKILL_DIR>/agent-browser" open https://app.example.com/login
"<SKILL_DIR>/agent-browser" snapshot -i
"<SKILL_DIR>/agent-browser" fill @e1 "username"
"<SKILL_DIR>/agent-browser" fill @e2 "password"
"<SKILL_DIR>/agent-browser" click @e3
"<SKILL_DIR>/agent-browser" wait --url "**/dashboard"
"<SKILL_DIR>/agent-browser" state save auth.json

# Later: load saved state
"<SKILL_DIR>/agent-browser" state load auth.json
"<SKILL_DIR>/agent-browser" open https://app.example.com/dashboard