agent-browser

What is agent-browser?

agent-browser is a browser automation CLI tool for AI agents developed by Vercel Labs. Written in native Rust, it directly controls the real Chrome browser based on the Chrome DevTools Protocol (CDP). Through a compact accessibility tree snapshot and ref reference mechanism, the tool provides LLMs with context-efficient web page interaction capabilities, supporting 50+ browser operation commands, multi-session isolation, cloud browser integration, and a real-time observability dashboard.

The main functions of agent-browser

• Snapshot Accessibility Tree Crawler : Outputs a compact, accessible tree structure for web pages, along with deterministic references such as `<ref> @e1` , for AI to accurately locate elements.@e2
• Ref-driven interaction : Perform operations such as clicking, filling, scrolling, and taking screenshots based on the Ref reference returned by the Snapshot, without repeatedly querying the DOM.
• 50+ browser operation commands : covering complete browser control capabilities including navigation, forms, mouse, keyboard, clipboard, network blocking, tab management, PDF export, etc.
• Semantic element location : Supports finding and manipulating elements using semantic methods such as ARIA roles, text content, tags, placeholders, and alt text.
• Multi-session isolation : Supports multiple instances of browser sessions running in parallel, each with its own independent cookies, storage, authentication, and navigation history.
• Network monitoring and interception : It can intercept, block, and mock network requests, and supports HAR recording and filtering of requests by resource type, status code, and method.
• Screenshot and Visual Diff : Supports full-page/viewport screenshots, numbered screenshots, pixel-level visual contrast, and Snapshot text diff.
• React / Web Vitals Introspection : Built-in React DevTools hooks that can output component trees, Fiber rendering records, Suspense boundaries, and performance metrics such as LCP/CLS/TTFB.
• Real-time streaming and dashboard : Push browser viewport content in real time via WebSocket, and the local dashboard provides activity streams, console output, and an AI Chat panel.
• AI Natural Language Control : Integrates with Vercel AI Gateway, agent-browser chatenabling browsers to perform complex tasks using natural language commands.

How to use agent-browser

• Install CLI : Perform npm install -g agent-browsera global installation of the native Rust binary.
• Initialize your browser : Run the command agent-browser installto automatically download Chrome for Testing, or reuse your existing Chrome/Brave installation.
• To open the target page : Launch agent-browser open example.comyour browser and navigate to the target URL.
• Get a page snapshot : Run the command agent-browser snapshot -ito output the accessibility tree of interactive elements and Ref references.
• Perform interactive operations : Use Ref references to perform operations such as agent-browser click @e2, , , .agent-browser fill @e3 "test@example.com"
• Screenshot or export : Execute agent-browser screenshot page.pngor agent-browser pdf report.pdfsave the result.
• Close session : Run agent-browser closethis command to end the current browser session.

The core advantages of agent-browser

• AI context efficiency is prioritized : compact text output significantly reduces LLM context usage compared to JSON or the full DOM, and is designed specifically for agent workflows.
• Real Browser Anti-Detection : Based on CDP, it controls the real Chrome browser instead of a headless browser, naturally carrying user profiles and reducing the probability of being blocked by risk control.
• Zero Node.js dependencies : Pure Rust native binary, no Playwright or Node runtime required, lightweight installation package, and extremely fast startup.
• Native cross-platform support : Provides natively compiled versions for macOS (ARM64/x64), Linux (ARM64/x64), and Windows (x64).
• Enterprise-level security mechanisms : Provides optional security layers such as encrypted authentication vault, content boundary markers, domain whitelists, action policy files, sensitive operation confirmation, and output length limits.
• Seamless cloud browser integration : Supports one-click access to cloud browser infrastructures such as Browserless, Browserbase, Browser Use, Kernel, and AWS AgentCore, and supports serverless deployment.

The project address for agent-browser

• Project website : https://agent-browser.dev/
• GitHub repository : https://github.com/vercel-labs/agent-browser