agent-browser

What is agent-browser?

agent-browser is a browser automation command-line tool purpose-built for AI agents and the coding assistants they run inside — including Claude Code, Cursor, GitHub Copilot, OpenAI Codex, and Google Gemini. It outputs a compact accessibility tree instead of raw HTML or JSON, drastically reducing token consumption while giving agents deterministic control over web pages through a ref-based element selection system. The tool is built on a client-daemon architecture: a native Rust CLI handles command parsing at near-instant speed, while a persistent Node.js daemon manages the underlying Playwright browser instance. With 50+ commands covering navigation, form interaction, screenshots, network inspection, and storage, agent-browser provides full-cycle browser control optimized for automated agent workflows across macOS, Linux, and Windows.

Key Features

• Ref-Based Element Selection

  Snapshot command returns a compact accessibility tree where every element carries a unique ref (e.g. @e1), enabling deterministic, re-query-free interaction without ambiguity.

• Token-Efficient Output

  Text-based accessibility tree output uses roughly 200–400 tokens per snapshot, compared to 3,000–5,000 tokens for full DOM output, keeping agent context windows lean.

• Native Rust CLI Performance

  Command parsing is handled by a native Rust binary that starts instantly, with a persistent Node.js daemon managing the Playwright browser instance in the background.

• 50+ Browser Commands

  Comprehensive command set covering page navigation, form filling, clicking, screenshots, network monitoring, and storage management for end-to-end browser control.

• Multi-Session Support

  Run multiple isolated browser instances simultaneously, each with independent authentication state, enabling parallel agent tasks or multi-account workflows.

• Cross-Platform Compatibility

  Native binaries available for macOS (ARM64 and x64), Linux (ARM64 and x64), and Windows (x64), with an npm fallback ensuring broad environment coverage.

Use Cases

• AI Coding Assistants : Agents inside Claude Code, Cursor, or GitHub Copilot can browse documentation, test web UIs, and fill forms as part of automated coding and debugging workflows.
• Web Scraping & Data Extraction : AI agents can navigate pages, capture snapshots, and extract structured information from websites with minimal token overhead per operation.
• Automated UI Testing : QA agents can script full browser interaction sequences — clicks, form submissions, screenshot comparisons — using deterministic refs rather than fragile CSS selectors.
• Multi-Step Web Automation : Complex workflows such as logging into services, filling multi-page forms, or monitoring network requests can be orchestrated through shell-command sequences any agent can issue.

FAQs