platonai/Browser4

🤖 Browser4 – AI‑native browser engine

What it is

• A Rust‑based browser automation platform that exposes a command‑line interface (browser4‑cli) and an HTTP‑backed backend. It is designed for both human users and autonomous AI agents that need to drive real browsers, extract data, and run large‑scale web‑crawls.

Core capabilities

Capability	How it works
Agent‑driven browsing	Agents (or humans) send commands like open, click, fill, goto, snapshot to a live Chromium instance via CDP. The CLI can run headless (default) or headed for visual debugging.
Zero‑token extraction	Uses X‑SQL (SQL‑style queries over the DOM) and CSS selectors to pull structured data from live pages or stored HTML snapshots. No LLM tokens are required for deterministic extraction.
Hybrid intelligence	LLM‑powered commands (extract, summarize, chat) can be mixed with the deterministic X‑SQL pipeline, letting you fall back to pure ML‑based clustering (WebMiner) when you want token‑free spreadsheets.
High‑throughput runtime	Coroutine‑safe, CDP‑native engine claims 100k‑200k page visits per machine per day. Supports swarm (distributed crawling) and batch/loop jobs for massive scale.
Programming‑Agent kernel	Over 50 built‑in coding.* tools give agents sandboxed shell/FS access, code scaffolding, validation, and self‑development, enabling agents to generate or modify Browser4 artifacts on the fly.
Extensible plugins & skills	Plugins can be added via the plugin command; skills (pre‑written automation scripts) live under skills/.

Typical workflow

1. Start a session – browser4-cli open https://example.com (add --headed to see the window).
2. Inspect – browser4-cli snapshot -i --boxes shows clickable elements with refs like e15.
3. Interact – click e15, fill e16 "text" --submit, wait --load networkidle.
4. Extract –
   • Simple field: htmlsnapshot get text "#price"
   • Structured list: htmlsnapshot query --sql @query.sql
   • LLM‑assisted: extract "find the product price" (requires an LLM key).
5. Scale – Use crawl, swarm create, or loop to process many URLs, then optionally run WebMiner to turn the collected HTML into spreadsheets without any LLM usage.

Installation

• Node: npm install -g browser4-cli && browser4-cli install
• One‑liner scripts for Windows PowerShell or Linux/macOS that download and bootstrap the native binary and runtime bundle.
• The backend is also published on Maven Central (ai.platon.pulsar:browser4-core) for Java/Scala users.

Key commands (human‑focused)

• Session control: open, attach, close, list, status, doctor
• Navigation: goto, go-back, reload
• Interaction: click, fill, type, press, hover, drag, wait
• Extraction: htmlsnapshot get, htmlsnapshot query, extract (LLM), eval --json
• Scaling: crawl, swarm create, swarm query, loop
• Reporting: screenshot, pdf, webminer (ML clustering → HTML report + Excel)

Why it matters for AI agents

• Provides a stable, deterministic API for agents to manipulate browsers, avoiding the brittleness of screen‑scraping alone.
• The zero‑token extraction path lets you keep costs down and privacy high when you only need structured data.
• The programming‑agent kernel gives agents the ability to write code, manage files, and evolve their own automation scripts without leaving the Browser4 ecosystem.

Resources

• Website: https://browser4.io
• Documentation: browser4-cli help … and the skills/ folder in the repo.
• Community: GitHub Issues, CI badge shows active testing, Maven Central and npm packages for easy consumption.

---

All details above are taken directly from the repository’s README; no additional features have been inferred.