agent-browser

What is agent-browser?

agent-browser is a high-performance, headless browser automation CLI specifically built for AI agents (like Claude Code, Cursor, or Gemini). Unlike traditional automation tools that return heavy JSON or HTML, it focuses on providing a compact accessibility tree (text-based) to maximize LLM context efficiency and reduce token costs.

Key Features

• Agent-first Design: Uses compact text output instead of bulky JSON/DOM, saving thousands of tokens per request.
• Ref-based Interaction: Assigns unique tags (e.g., @e1, @e2) to elements in a snapshot, allowing AI to interact with specific elements deterministically.
• Native Rust Performance: Built entirely in Rust for instant command parsing and execution.
• Session Management: Supports multiple isolated browser instances with separate authentication states.
• Comprehensive Command Set: Includes over 50 commands covering navigation, form filling, screenshots, network monitoring, and storage.

Use Cases

• AI Agent Integration: Providing LLMs with a "browser-as-a-tool" to perform web research or data entry.
• Web Automation: Running scripts for automated testing or repetitive web tasks via shell commands.
• Context-Constrained Environments: Navigating complex web pages within limited token windows.
• Cross-Platform Automation: Deploying automation tasks across macOS, Linux, and Windows using native binaries.

FAQ / Why use "Refs"?

The tool utilizes a "Ref-based" snapshot system because:

1. Efficiency: Text snapshots use ~200-400 tokens, compared to ~3000-5000 for a full DOM.
2. Precision: Refs point to exact elements, eliminating the need for the AI to re-query the DOM or guess selectors.
3. Speed: No DOM re-querying is required between commands.