Browser Automation Skill (agent-browser)

This skill provides full browser automation via the agent-browser CLI tool.

Documentation: https://agent-browser.dev | GitHub: https://github.com/vercel-labs/agent-browser

When to Use This Skill

Use agent-browser when you need to:

•Navigate websites interactively
•Click buttons and links
•Fill forms
•Take screenshots or PDFs
•Extract data from pages
•Explore unknown UIs step by step
•Perform lightweight browser automation

Mandatory Session Workflow

For EVERY agent-browser session:

•GENERATE a random <sessionId> (e.g., abc123)
•START with agent-browser --session <sessionId> open <url>
•ALL subsequent commands must include --session <sessionId>
•CLEANUP (REQUIRED): Every session MUST end with agent-browser --session <sessionId> close

⚠️ Never leave browser sessions open — always close them to prevent resource leaks.

Core Workflow Examples

Step 1 — Open a page (Start Session)

bash

agent-browser --session abc123 open https://example.com

Step 2 — Snapshot the page

bash

agent-browser --session abc123 snapshot -i

This returns interactive elements with references:

code

textbox "Email"        [ref=e1]
textbox "Password"     [ref=e2]
button  "Submit"       [ref=e3]

Step 3 — Interact using refs

bash

agent-browser --session abc123 fill @e1 "user@example.com"
agent-browser --session abc123 fill @e2 "password123"
agent-browser --session abc123 click @e3

Step 4 — Re-snapshot after navigation

bash

agent-browser --session abc123 snapshot -i

Step 5 — Close Session (REQUIRED)

bash

agent-browser --session abc123 close

Navigation Commands

bash

agent-browser --session <sessionId> open <url>
agent-browser --session <sessionId> back
agent-browser --session <sessionId> forward
agent-browser --session <sessionId> reload
agent-browser --session <sessionId> close

Snapshot (Page Analysis)

bash

agent-browser --session <sessionId> snapshot
agent-browser --session <sessionId> snapshot -i          # Interactive elements only
agent-browser --session <sessionId> snapshot -c          # Compact output
agent-browser --session <sessionId> snapshot -d 3        # Depth limit
agent-browser --session <sessionId> snapshot -s "#main"  # Scope to selector

Interactions (Using Refs from Snapshot)

bash

agent-browser --session <sessionId> click @e1
agent-browser --session <sessionId> dblclick @e1
agent-browser --session <sessionId> hover @e1
agent-browser --session <sessionId> focus @e1
agent-browser --session <sessionId> fill @e2 "text"
agent-browser --session <sessionId> type @e2 "text"
agent-browser --session <sessionId> press Enter
agent-browser --session <sessionId> check @e1
agent-browser --session <sessionId> uncheck @e1
agent-browser --session <sessionId> select @e1 "value"
agent-browser --session <sessionId> scroll down 500
agent-browser --session <sessionId> scrollintoview @e1
agent-browser --session <sessionId> drag @e1 @e2
agent-browser --session <sessionId> upload @e1 file.pdf

Getting Information

bash

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

Screenshots & PDFs

bash

agent-browser --session <sessionId> screenshot
agent-browser --session <sessionId> screenshot page.png
agent-browser --session <sessionId> screenshot --full
agent-browser --session <sessionId> pdf output.pdf

Waiting

bash

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

Complete Example: Form Submission

bash

# Start session
agent-browser --session myform open https://example.com/form

# Analyze the page
agent-browser --session myform snapshot -i

# Fill and submit
agent-browser --session myform fill @e1 "user@example.com"
agent-browser --session myform fill @e2 "password123"
agent-browser --session myform click @e3

# Wait for navigation and re-analyze
agent-browser --session myform wait --load networkidle
agent-browser --session myform snapshot -i

# ALWAYS close the session
agent-browser --session myform close

Quick Reference

Action	Command
Open URL	`agent-browser --session ID open URL`
Get page structure	`agent-browser --session ID snapshot -i`
Click element	`agent-browser --session ID click @ref`
Fill input	`agent-browser --session ID fill @ref "text"`
Take screenshot	`agent-browser --session ID screenshot file.png`
Close session	`agent-browser --session ID close`

My Principle

Explore → Snapshot → Interact → Close. Always use snapshots to understand the page structure before interacting. Always close sessions when done. Keep automation interactive and incremental.