platonai/Browser4

🤖 Browser4 – What It Is

Browser4 is an AI‑native browser automation engine. It lets large‑language‑model agents (or humans) drive a real Chromium‑based browser via a fast Rust‑written CLI (browser4‑cli). The system combines traditional web‑automation (click, fill, scroll, screenshot) with zero‑token extraction – deterministic data extraction using X‑SQL and CSS selectors, plus optional LLM‑powered extraction. It also ships a WebMiner pipeline that clusters HTML pages into spreadsheets without any LLM calls.

---

Key Capabilities (from the README)

Capability	What it means
Agent Browser	Agents can open a browser, interact (click, fill, snapshot), run batch jobs, and loop over pages.
Zero‑Token Extraction	Use X‑SQL + CSS selectors or WebMiner ML clustering to turn pages into structured tables without spending LLM tokens.
Hybrid Intelligence	Mix LLM extraction, ML clustering, and a reusable selector store.
High‑Performance Runtime	Coroutine‑safe, CDP‑native engine designed for 100k‑200k page visits per machine per day, supporting swarm/crawl scale‑out.
Enterprise‑Scale Automation	Swarm crawling, batch/loop jobs, stateful sessions, plugins, runtime skills, browser extensions, and an HTTP‑based MCP.
Programming‑Agent Kernel	50+ coding.* tools for agents to generate, validate, and self‑develop Browser4 artifacts.

---

How It Works (Architecture Overview)

• Backend – a Java/Kotlin runtime (published on Maven Central as browser4‑core) that talks to Chrome/Edge via the Chrome DevTools Protocol (CDP). It manages sessions, plugins, skills, and a Pulsar SDK for agent state.
• CLI (browser4‑cli) – a Rust wrapper that sends HTTP/MCP commands to the backend. It provides a shell‑like interface (open, click, snapshot, htmlsnapshot, crawl, swarm, …) usable by humans or AI agents.
• Modules – include a programming‑agent kernel (browser4‑coding), a test fixture server (MockSite), and integration with the WebMiner ML clustering tool.

---

Typical Use Cases

1. Autonomous agents – Prompt an LLM (e.g., Claude, DeepSeek) with a short skill file, have it call browser4‑cli to navigate a site, fill forms, and extract data.
2. Zero‑token data pipelines – Crawl a list of URLs, run htmlsnapshot query --sql … or webminer to produce Excel/HTML reports without any LLM cost.
3. Large‑scale crawling – Use swarm create + swarm query to distribute tens of thousands of page visits across machines.
4. Human‑in‑the‑loop testing – Open a headed browser, inspect element refs (snapshot -i --boxes), then script further actions.

---

Getting Started (Quick‑Start Steps)

1. Install – npm install -g browser4-cli then browser4-cli install (or run the provided curl/PowerShell script).
2. Open a browser – browser4-cli open --headed https://example.com.
3. Inspect – browser4-cli snapshot --boxes to see element refs like e15.
4. Interact – browser4-cli click e15, browser4-cli fill e16 "text" --submit.
5. Extract – browser4-cli get text "h1" or browser4-cli htmlsnapshot query --sql @query.sql.
6. Scale – Use crawl, swarm, or loop commands for batch or periodic jobs.

---

Ecosystem & Integration

• DeepSeek Harness – A plugin (dsh-browser4) lets the dsh agent framework call Browser4 directly.
• WebMiner – Runs locally (SMILE ML library) to cluster HTML into spreadsheets; commercial version uses Apache Spark for >1k pages.
• Plugins & Skills – Extend functionality via the built‑in plugin system; view with browser4-cli plugin-list.
• LLM Providers – Supports DeepSeek, OpenRouter, Volcengine, OpenAI‑compatible, and Aliyun Qwen via environment variables.

---

Documentation & Support

• Website – https://browser4.io
• GitHub – Issues, releases, CI badge, Maven Central artifact.
• CLI Help – browser4-cli --help and browser4-cli help <command>.
• Status Dashboard – http://127.0.0.1:18182/status shows live health, sessions, swarm tasks, etc.
• Community – Support links listed under “Support & Community” in the repo.

---

License

Apache‑2.0.

---

Bottom line: Browser4 is a full‑stack, high‑performance browser automation platform designed for AI agents and large‑scale web data extraction, offering both traditional Selenium‑style interactions and token‑free, ML‑driven structuring of web content.