agentbrowser

name	AgentBrowser
description	Browser automation for development testing, web scraping, and UI verification. USE WHEN you need to interact with web pages, verify UI changes, take screenshots, or extract data from websites.
allowed-tools	Bash(agent-browser:*)

You control a headless Chromium browser for web automation tasks. This skill provides CLI commands to navigate, interact with, and inspect web pages. The daemon starts automatically on first use — no explicit start command needed.

When Invoked

1. Determine Action: Based on the user's request:

   • Need to view a page? → Use open then snapshot
   • Need to interact? → Use click, fill, or eval
   • Need visual verification? → Use screenshot
   • Need to wait for content? → Use wait

2. Execute Commands: Use the appropriate CLI commands via Bash tool

3. Report Results: Summarize what you found or accomplished

Session Management

Sessions persist browser state (cookies, localStorage) across commands. Use --session before the subcommand to specify a named session.

• Default session: If --session is omitted, the "default" session is used
• Multiple sessions: Different agents can use different session IDs
• Auto-cleanup: Sessions auto-terminate after 30 minutes of inactivity

Command Reference

Command	Usage	Description
open	agent-browser --session <id> open <url>	Navigate to URL
snapshot	agent-browser --session <id> snapshot -i	Get interactive elements with refs
click	agent-browser --session <id> click <ref|selector>	Click element
fill	agent-browser --session <id> fill <ref|selector> <text>	Clear and fill input
screenshot	agent-browser --session <id> screenshot [path]	Take screenshot
eval	agent-browser --session <id> eval <script>	Run JavaScript
wait	agent-browser --session <id> wait <ref|selector>	Wait for element
close	agent-browser --session <id> close	Close browser session

See commands.md for the full command set.

Workflow Routing

Workflow	Trigger Words	When to Use
BrowseWeb	"browse", "open", "navigate", "check", "verify", "screenshot", "scrape"	User wants to interact with or inspect a web page

Reference

• Commands — Full command reference with all options
• Snapshot Refs — The @e1, @e2 locator system
• Session Management — Named sessions, state persistence

Tips

• Always use snapshot -i after navigation to understand page structure
• Use the locators (@e1, @e2) from snapshot for click/fill selectors — they are the most reliable
• Re-snapshot after navigation or DOM changes; refs are invalidated on page changes
• Keep the same --session ID across related commands to maintain state
• Use screenshot --annotate for visual context alongside text snapshots