A browser CLI that treats AI as the primary user

What it does

agent-browser is a native Rust CLI that launches and controls Chrome—detecting existing Chrome, Brave, Playwright, or Puppeteer installations automatically—through a terse command interface. It covers the standard browser automation primitives like clicking, filling, scrolling, screenshots, and PDF export, but also returns an accessibility tree with reference IDs via snapshot so LLMs can target elements without guessing at CSS selectors. The daemon runs without Playwright or Node.js.

The interesting bit

The tool treats the machine as the primary user. The snapshot command emits stable refs such as @e2, and the chat command offers natural-language browser control in both single-shot and interactive REPL modes. A batch mode and JSON-over-stdin interface make it clear this is designed to be wired into larger agent loops rather than used for one-off human scripting.

Key highlights

• Distributed as a native Rust binary via npm, Homebrew, and Cargo; self-updates and can self-install Chrome from Google’s “Chrome for Testing” channel.
• Semantic locators (find role button click --name "Submit") alongside traditional CSS selectors.
• Network interception, HAR recording, request filtering, and mock responses for simulating different agent environments.
• Clipboard, mouse control, geolocation, device emulation, offline mode, and annotated screenshots with numbered element labels.
• eval for arbitrary JavaScript execution and a WebSocket streaming mode for runtime telemetry.

Caveats

• The README catalogs an extensive command surface but does not explain how the chat command processes natural language (local model, cloud API, etc.).
• Building from source requires Node.js 24+, pnpm 11+, and Rust—a heavy toolchain for a tool whose main selling point is avoiding Node.js dependencies.

Verdict

A solid choice if you are building autonomous agents or data pipelines that need to drive a real browser without a Node.js automation stack. Less compelling if you want a human-centric testing framework with built-in assertions and test runners.