agent-browser: Rust-based browser CLI directly controlled by AI agents from the shell

A terminal-based tool that allows AI agents to open web pages, read content as an accessibility tree, and perform actions such as clicking or filling specific elements. A native CLI written in Rust controls the browser via the Chrome DevTools Protocol through a background daemon. Each command is handled as a short shell call, and the browser session is maintained between commands, enabling continuous tasks.

It is designed to assign references (refs) like @e1 to all page elements, allowing agents to execute actions through these references. Clicks on hidden elements, disabled buttons, or elements outside the viewport are reported as failures rather than successes, prompting the agent to recheck the state and retry. CSS selectors, XPath, and semantic locators are also supported, allowing elements to be specified in various ways.

It uses Chrome for Testing by default but can connect to a running Chrome instance or select the Lightpanda or Obscura engines. It also supports running Mobile Safari on the iOS simulator and integration with cloud browser providers such as Browserbase and Browserless. It includes a stdio-based MCP server for MCP clients, making it easy to integrate into existing agent infrastructure.

It provides skills for agents to help automatically load instructions matching the CLI version. It distributes native binaries supporting ARM64 and x64 architectures on macOS, Linux, and Windows, and uses an npm package to automatically select the binary suitable for the host environment. It is released under the Apache-2.0 license and is suitable for developers who want to perform web tasks with simple shell commands instead of writing complex browser automation scripts.