agent-browser: Fast Rust CLI for AI Browser Automation

Why a New Tool for Browser Automation?

AI agents are increasingly required to interact with real‑world web pages: testing UI flows, scraping dynamic content, or crawling e‑commerce sites. Existing tools like Selenium or Playwright are fantastic libraries, but they demand a learning curve, setup overhead, and often a heavy runtime. If an AI system (Claude, GPT‑4, Gemini, etc.) were to control a browser, it would be convenient to expose that capability via a lightweight, command‑line interface that can be called from any language.

Enter agent-browser – a single executable that ships a fast Rust binary on macOS, Linux, and Windows, with automatic fall‑back to Node.js when the native binary is unavailable. It talks to a Node.js daemon that owns a Playwright instance, giving you all the power of Playwright (Chromium, Firefox, WebKit) while keeping the CLI fast and deterministic.

Quick Start

# Install globally via npm
npm install -g agent-browser
# Grab the bundled Chromium
agent-browser install
# Open a site
agent-browser open https://example.com
# Take a reference‑based snapshot
agent-browser snapshot --json
# Click a button by ref
agent-browser click @e2
# Take a screenshot
agent-browser screenshot page.png
# Close it
agent-browser close

Installation Options

• Rust Native – pnpm build:native requires Rust 1.75+. Gives the best startup time.
• Node.js Fallback – Works everywhere but slightly slower.
• Linux Dependencies – agent-browser install --with-deps will install Playwright’s system‑level dependencies automatically.

Features & Commands

Architecture Snapshot

• Rust CLI – Parses commands, builds a JSON DSL, and sends them to the daemon.
• Node.js Daemon – Launches and manages the Playwright browser; keeps the process alive between invocations.
• Client‑Daemon Design – One daemon per system, so repeated commands don’t pay the cost of launching a new Chromium instance.

How AI Agents Use It

1. Open a page.
2. Snapshot with --json. The result contains a structured tree (role, name, ref).
3. The AI reasoner parses the tree to find the target ref (e.g., @e3 for an email box).
4. Act using CLI commands (click @e3, fill @e4 "[email protected]").
5. Re‑snapshot if the page changed, then repeat.

The CLI’s JSON output is tiny (~2‑3 KB) and ideal for LLMs with limited prompt tokens.

Use Cases

Example: Building a Mini‑Scraper in 10 Lines

#!/usr/bin/env node
const { execSync } = require('child_process');

execSync('agent-browser install');
execSync('agent-browser open https://books.toscrape.com');
const data = execSync('agent-browser snapshot -i --json').toString();
const snapshot = JSON.parse(data).data.snapshot;
console.log('Page has books:', snapshot.children.filter(c=>c.role==='link').length);
execSync('agent-browser close');

Run it with node scrape.js and you’ll get a quick count of book links, all done via the CLI.

Performance & Reliability

• Native Rust binary boots in < 50 ms on macOS – compared to Node‑only counterparts that take 200–300 ms.
• The daemon remains resident after the first run, so subsequent commands are virtually instantaneous.
• Works on all major platforms; the binary is distributed via GitHub releases.

Getting Involved

• Contributions welcome on GitHub: create a fork, submit a PR, or open issues for bugs.
• Want more semantics? Add new find actions or support more browser engines.
• Documentation is living – check the docs/ folder and the CLI --help output for the latest commands.

Final Thoughts

agent‑browser brings web automation to the fingertips of any AI system or developer with a single command. Its blend of Rust speed, Playwright versatility, and AI‑friendly reference snapshots makes it a compelling tool for testing, scraping, and building agent‑centric workflows. Whether you’re automating a CI pipeline or pairing a large‑language‑model with live browsing, agent‑browser gives you a lightweight, deterministic, and easy‑to‑program interface that feels at home in a shell or a serverless function.

Happy automating!