Agent Browser

Overview

Agent Browser is a fast, native Rust CLI for browser automation designed specifically for AI agents. It provides Chrome/Chromium control via the Chrome DevTools Protocol (CDP) with accessibility-tree snapshots and compact @eN element references, enabling reliable interaction with web pages without Playwright or Puppeteer dependencies.

Released in mid-2026, Agent Browser quickly gained traction as the go-to browser automation layer for coding agents like Claude Code, Cursor, Codex, Windsurf, and Gemini CLI. By the end of June 2026, it had achieved 5,000+ weekly active users and was integrated into the Claude Code skills ecosystem as a first-class plugin.

Features

• Native Rust CLI: Sub-second startup, no Node.js runtime overhead
• Chrome DevTools Protocol: Direct CDP connection for fast, reliable automation
• Accessibility-tree snapshots: Semantic DOM understanding via browser accessibility tree
• Compact element refs: @eN notation for precise element targeting
• No Playwright/Puppeteer: Smaller footprint, faster execution
• Session management: Persistent sessions, authentication vault, state persistence
• Video recording: Built-in recording for debugging and audit trails
• Multi-agent compatible: Works with Claude Code, Cursor, Codex, Continue, Windsurf, Gemini CLI
• Electron app support: Specialized skills for automating desktop apps (VS Code, Slack, Discord, Figma)
• Cloud browser support: Integrations with Vercel Sandbox and AWS Bedrock AgentCore

Installation

npm i -g agent-browser && agent-browser install

Requires Chrome or Chromium installed on the system.

Quick Start

# Start an interactive session
agent-browser

# Launch a headless session
agent-browser --headless

# Load a specialized skill
agent-browser skills get core
agent-browser skills get electron  # for desktop apps
agent-browser skills get slack     # for Slack automation

Core Concepts

Element References (@eN)

Agent Browser uses compact @eN references derived from the accessibility tree:

agent-browser click @e3       # click element with ref @e3
agent-browser type @e5 hello  # type text into element @e5

Accessibility-tree Snapshots

Instead of parsing raw DOM, Agent Browser queries the browser's accessibility tree, which is already optimized for screen readers and contains semantic meaning (button, link, heading, etc.).

Session Management

agent-browser session list      # list active sessions
agent-browser session save foo  # save current state as "foo"
agent-browser session load foo  # restore session

Advanced Features

Specialized Skills

Agent Browser ships with domain-specific automation skills:

Skill	Use Case
electron	Automating VS Code, Slack, Discord, Figma
slack	Workspace automation (search, send, reply)
dogfood	Exploratory testing, QA, bug hunts
vercel-sandbox	Running inside Vercel microVMs
agentcore	AWS Bedrock cloud browsers

Observability Dashboard

Agent Browser includes an observability dashboard running on port 4848, providing:

• Real-time session status
• Traffic streams
• Video playback
• Session history

Accessible at https://dashboard.agent-browser.localhost.

Video Recording

Built-in video recording captures browser sessions for debugging, auditing, and training data:

agent-browser --record recordings/session.mp4

Examples

Web Scraping

agent-browser "Open https://example.com, take a screenshot, extract all headings"

Form Automation

agent-browser "Go to login page, fill email and password fields, submit form"

Desktop App Automation (Electron)

agent-browser skills get electron
agent-browser "Open VS Code, open file settings.json, add new setting, save"

Slack Automation

agent-browser skills get slack
agent-browser "Search Slack for messages containing 'deployment', summarize results"

Pros

• ✅ Fast native Rust CLI with sub-second startup
• ✅ No Playwright/Puppeteer dependency — smaller footprint
• ✅ Accessibility-tree snapshots for semantic element understanding
• ✅ Works with any AI agent (Claude Code, Cursor, Codex, Windsurf, etc.)
• ✅ Specialized skills for domain-specific automation (Electron, Slack, etc.)
• ✅ Built-in video recording and observability dashboard
• ✅ Session persistence and authentication vault

Cons

• ❌ Requires Chrome/Chromium installation
• ❌ Relatively new (as of July 2026), ecosystem still growing
• ❌ Documentation is CLI-driven (skills get core), may feel less discoverable
• ❌ Windows support may require additional setup for some features

When to Use

• Web automation for AI agents: When you need an agent to interact with web pages reliably
• Testing and QA: Automating exploratory testing, bug reproduction, or UI regression checks
• Desktop app automation: Controlling Electron apps like VS Code, Slack, or Figma
• Cloud browser scenarios: Running browser automation in Vercel Sandbox or AWS Bedrock
• Form automation: Filling out forms, logging in, navigating multi-step flows

Integration with Claude Code

Agent Browser is available as a first-class skill in Claude Code. When triggered by user requests involving browser interaction, Claude Code automatically loads the Agent Browser skill and executes commands through the CLI.

# Claude Code automatically handles:
# - Launching browser sessions
# - Taking accessibility-tree snapshots
# - Interacting with elements via @eN refs
# - Recording sessions for debugging

Resources

• GitHub Repository
• npm Package
• Claude Code Skills Documentation
• Chrome DevTools Protocol

---

Published: July 3, 2026