agent-browser（Agent 浏览器自动化）

agent-browser Skill

Browser automation that actually works for AI agents. Built by Vercel Labs specifically for LLM-driven workflows.

Why This Works Better Than Alternatives

1. Deterministic Refs (The Game-Changer)

Problem with traditional tools:

• CSS selectors break when websites change
• XPath is brittle and unreadable
• Coordinate-based clicking fails on responsive layouts
• Vision-based approaches are slow and expensive

The agent-browser solution:

# 1. Get snapshot with stable refs
agent-browser snapshot -i --json
# Output: - button "Submit" [ref=e2]

# 2. Use that ref forever — it points to the EXACT element
agent-browser click @e2

• Refs are deterministic — @e2 always points to the same element from your snapshot
• No DOM re-query — direct reference is faster and more reliable
• AI-optimized — LLMs parse the accessibility tree naturally, not CSS soup

2. Accessibility Trees > Screenshots/HTML

Traditional tools give you raw HTML (noisy) or screenshots (require vision models).

agent-browser gives you the accessibility tree — a clean, semantic representation of what a human (or screen reader) would perceive:

- heading "Billing" [level=1]
- link "Make a payment" [ref=e10]
- button "Submit" [ref=e2]
- textbox "Email" [ref=e3]

• Semantic roles (button, link, textbox, heading)
• Human-readable labels
• Hierarchical structure
• Perfect for LLM comprehension

3. Built for AI Agents

4. Fast Architecture

• Rust CLI — Native binary, instant command parsing
• Node.js Daemon — Browser stays warm between commands
• First command: ~2s (daemon startup)
• Subsequent commands: ~100ms

Prerequisites

npm install -g agent-browser
agent-browser install  # Download Chromium (~30s)

Core AI Workflow

The workflow designed for LLM agents:

# Step 1: Navigate
agent-browser open https://example.com

# Step 2: Get structured snapshot (the AI "sees" the page)
agent-browser snapshot -i --json

# Step 3: AI picks refs from JSON, execute actions
agent-browser click @e2
agent-browser fill @e3 "test@example.com"

# Step 4: Re-snapshot after changes (state verification)
agent-browser snapshot -i --json

# Step 5: Done
agent-browser close

Commands

Navigation

agent-browser open example.com
agent-browser open example.com --json            # JSON response
agent-browser open example.com --headed          # Visible browser

Snapshot (The Killer Feature)

agent-browser snapshot                           # Full accessibility tree
agent-browser snapshot -i                        # Interactive only (faster)
agent-browser snapshot -i --json                 # JSON for AI parsing
agent-browser snapshot -i -c -d 5 --json         # Compact, depth-limited

Interaction (Using Deterministic Refs)

agent-browser click @e2                          # Click element @e2
agent-browser fill @e3 "text"                    # Fill and clear
agent-browser type @e3 "text"                    # Type without clearing
agent-browser press Enter                        # Press key
agent-browser hover @e4                          # Hover

State Verification

agent-browser get text @e1                       # Get element text
agent-browser get url                            # Current URL
agent-browser is visible @e2                     # Check visibility

Session Management

agent-browser --session login open site.com      # Isolated session
agent-browser --profile ~/.myprofile open site   # Persistent cookies
agent-browser close                              # Clean up

Selector Strategies (Ranked by Reliability)

1. Refs (Best - Use These)

# From snapshot output — deterministic and stable
agent-browser click @e2
agent-browser fill @e3 "text"

2. Semantic Locators (Good)

agent-browser find role button click --name "Submit"
agent-browser find label "Email" fill "test@test.com"

3. CSS Selectors (Okay for static sites)

agent-browser click "#submit"
agent-browser click ".btn-primary"

4. Text/XPath (Last resort)

agent-browser click "text=Submit"
agent-browser click "xpath=//button[1]"

Snapshot Options

Control what the AI "sees":

Recommended AI command:

agent-browser snapshot -i -c --json

Options

Example: Complete Login Flow

# Start
agent-browser open https://portal.aeronetpr.com

# Get page structure
SNAPSHOT=$(agent-browser snapshot -i --json)
# AI parses JSON: sees textbox @e1 (Username), textbox @e2 (Password), button @e3 (Login)

# Execute login
agent-browser fill @e1 "username"
agent-browser fill @e2 "password"
agent-browser click @e3

# Verify success (wait for navigation, re-snapshot)
sleep 2
agent-browser snapshot -i --json

# Done
agent-browser close

Tips for AI Agents

1. Always use --json — Structured output is easier to parse than text
2. Use -i flag — Interactive-only snapshots are smaller, faster, cleaner
3. Re-snapshot after actions — Verify state changed as expected
4. Trust refs over selectors — @e2 from snapshot > #id that might change
5. Use semantic locators when refs expire — find role button click is robust
6. Session persistence — One open, many commands, one close

Comparison to Other Tools

When to Use This Skill

Use agent-browser when:

• Automating multi-step web workflows
• Filling complex forms
• Need reliable, repeatable automation
• Working with dynamic/modern web apps
• Cost matters (no vision API calls)

Use OpenClaw's built-in browser tool when:

• Simple single-page checks
• Quick screenshot needed
• Already authenticated session in Chrome

Resources

• Vercel Labs repo: https://github.com/vercel-labs/agent-browser
• This skill repo: https://github.com/clawdbrunner/skill-agent-browser