agent-browser is a command-line interface (CLI) tool designed to facilitate browser automation specifically for artificial intelligence (AI) agents. Its core objective is to provide a highly efficient, compact, and deterministic method for AI agents to interact with web browsers by executing shell commands.

The core methodology of agent-browser revolves around a "ref-based" interaction model and a client-daemon architecture. When an AI agent needs to perceive the current state of a web page, the snapshot command is invoked. This command generates a compact, text-based representation of the page's accessibility tree, where each interactive or notable element is assigned a unique, stable reference identifier, denoted as @ref (e.g., @e1, @e2). This accessibility tree output is significantly more context-efficient than a full Document Object Model (DOM), typically consuming 200-400 tokens compared to 3000-5000 tokens for a complete DOM representation. The use of these refs ensures deterministic element selection, as a ref points to an exact element identified in the preceding snapshot, eliminating the need for subsequent DOM re-querying and guaranteeing consistent interaction. This text-based, ref-annotated output is specifically optimized for natural parsing by large language models (LLMs), making it AI-friendly.

The technical architecture of agent-browser is a client-daemon model engineered for optimal performance and persistence. The client-side component is a fast, native Rust CLI responsible for parsing commands and communicating with the daemon. The daemon, which manages the browser instance, exists in two primary forms:

1. Node.js Daemon (default): This daemon leverages Playwright to control browser instances. It provides robust, cross-browser automation capabilities.
2. Native Daemon (experimental): Written entirely in Rust, this daemon communicates directly with the browser using the Chrome DevTools Protocol (CDP), bypassing Node.js for potentially even lower overhead.

The daemon starts automatically upon the first command invocation and persists across subsequent commands, ensuring isolated and stateful browser sessions without repeated startup overhead.

Key features contributing to its suitability for AI agents include:

• Compact Text Output: Minimizes token usage for AI context windows.
• Ref-based Interaction: Provides deterministic, efficient, and AI-friendly element selection.
• High Performance: Achieved through a native Rust CLI and efficient client-daemon communication, with refs eliminating costly DOM re-queries.
• Comprehensive Command Set: Offers over 50 commands covering navigation, form interactions, screenshots, network monitoring, and storage management.
• Session Management: Supports multiple isolated browser instances, each with separate authentication contexts.
• Cross-platform Compatibility: Native Rust binaries are provided for macOS (ARM64, x64), Linux (ARM64, x64), and Windows (x64).

This design allows AI agents to perform complex browser automation tasks by chaining simple shell commands, such as agent-browser open example.com, agent-browser snapshot -i, and agent-browser click @e2, leveraging the deterministic and compact feedback loop provided by the ref system.