EyeBrowse MCP Server

은밀한 LLM 구동 브라우저 엔진 — Python 라이브러리 + 은밀 Chromium에서 전체 Chrome DevTools Protocol을 갖춘 85개 도구 MCP 서버

문서

👁️ EyeBrowse

A stealthy, LLM-drivable browser engine — one codebase, two faces.

A Python library and an MCP server for driving a real, hard-to-detect browser, so legitimate automation isn't false-flagged or IP-banned by Cloudflare, DataDome, Akamai, or PerimeterX. Built on CloakBrowser — a stealth Chromium (Chrome/146) that's a Playwright drop-in — so EyeBrowse gets the full Chrome DevTools Protocol: trusted cursorless clicks, deep network inspection, MHTML, PDF, and native video.

CI PyPI Python 3.12 License: MIT
MCP tools Engine: CloakBrowser Code style: Ruff PRs welcome

EyeBrowse — an AI agent driving a stealth browser past Cloudflare, over MCP

▶ Full-quality MP4: docs/demo.mp4 — an AI agent drives EyeBrowse over MCP: clears a Cloudflare check, then reads real docs (asyncio · httpx · MDN).


Why EyeBrowse?

  • 🥷 Stealth by default — engine-level fingerprint spoofing (geoip + humanize on out of the box, novel fingerprint per launch); navigator.webdriver masked; viewport auto-sized to the spoofed screen. No puppeteer-extra band-aids — the anti-detection is compiled into the browser.
  • 🤖 Built for LLMs — pages are read as an ARIA tree with [ref=…] handles; the model acts by ref (click/type/hover), not by brittle CSS or raw pixels. Cross-origin iframes, shadow DOM, popups — handled.
  • Chrome DevTools Protocoltrusted, cursorless clicks by node ref (Input.dispatchMouseEvent), raw Network/Performance/Emulation access, MHTML snapshots, PDF export, and native video — all reachable as tools.
  • 🧰 Library and MCP from one codebase — a clean Python API (EyeBrowse + Session), mirrored 1:1 by a thin MCP server (85 browser_* tools) for Claude Code and any MCP client.
  • 🪟 Never boxed in — the curated high-level API doesn't hide Playwright: reach session.page / .context / .browser for anything it doesn't wrap.
  • 🔋 Batteries included — multi-session, proxy + identity rotation, API-mode captcha solvers, native video, full HAR capture, and clean-markdown extraction.

Scope. EyeBrowse is a low-level browser engine — it holds no workflow logic. Consumers decide what to do; the engine provides what's possible.

Contents

Quickstart · Install · Features · Compare · Library · MCP · Proxy & identity · Extraction · Recording · How it works · Caveats · Tools · License

Quickstart

pip install eyebrowse
# The stealth-Chromium binary downloads automatically on first launch — nothing else to run.
import asyncio
from eyebrowse import EyeBrowse

async def main():
    eb = EyeBrowse()                          # stealth defaults: geoip · humanize
    async with eb.session() as s:
        await s.navigate("https://example.com")
        print(await s.snapshot())             # ARIA tree with [ref=...] handles
        await s.click("e6")                   # act on a ref from the snapshot
    await eb.aclose()

asyncio.run(main())

…or wire it into Claude Code (or any MCP client) — see Use over MCP.

Install

From PyPI

pip install eyebrowse                 # or: uv pip install eyebrowse
# CloakBrowser fetches its Chromium binary lazily on first launch — nothing to run.
pip install "eyebrowse[extract]"     # optional: + Crawl4AI markdown extraction (heavier)

From source (development)

git clone https://github.com/Evil-Bane/eyebrowse && cd eyebrowse
uv sync                              # core engine  (add --extra extract for Crawl4AI)
cp .env.example .env                 # only if you use a proxy / captcha keys

Python 3.12 (pinned <3.13). Engine: cloakbrowser>=0.3 (stealth Chromium, Chrome/146), on playwright 1.60 and mcp 1.27.

Features

🥷 StealthCloakBrowser's patched-Chromium fingerprint spoofing (novel --fingerprint per launch); geoip + humanize by default; webdriver masked; viewport matched to the spoofed screen.
🤖 LLM interactionaria_snapshot(mode="ai") → ARIA tree + [ref] handles; click / type / hover / select / drag / file-upload / dialogs / keyboard; coordinate mouse too.
CDPtrusted cursorless click by ref, raw CDP (Network / Performance / Emulation), MHTML capture, PDF export.
🪟 Frames & DOMcross-origin iframe routing by ref, shadow-DOM piercing, popup/new-tab switching, evaluate inside any frame.
🗂 Multi-sessionindependent stealth sessions, each with its own context / identity / proxy.
🌐 Networkinspect requests/responses (incl. XHR/fetch bodies & WebSocket frames), block URLs, mock responses, go offline, full HAR export.
💾 Statecookies, localStorage & sessionStorage (CRUD), storage_state save/reload.
🪪 Identity rotationfresh fingerprint + isolated profile + paired proxy; pluggable residential ProxyProvider.
🧩 Captchapluggable API-mode solvers (CapSolver / 2Captcha / CapMonster / NextCaptcha) + TOTP — no browser extension.
📄 ExtractionCrawl4AI raw: feed → clean, token-efficient markdown (no LLM, no API keys).
🎥 Capturescreenshots, Playwright tracing, and native video (.webm).
Verify & debugassertions, element highlighting, locator generation, geolocation/header emulation.

Full per-tool reference: docs/TOOLS.md (85 tools across 18 groups).

How EyeBrowse compares

EyeBrowsePlaywright MCPbrowser-useplaywright-stealth
Anti-detection compiled into the browser⚠️ JS patches
LLM-native ARIA [ref] interaction model
Ships an MCP server✅ (85 tools)⚠️ partial
One codebase: Python library and MCPMCP-onlylib-onlylib-only
Full CDP (trusted clicks · network · MHTML · PDF · video)⚠️ partial⚠️ partial
Captcha (API-mode) + TOTP
Proxy + identity rotation built in⚠️ partial
Cross-origin iframes · shadow DOM · popups⚠️ partialn/a

Fair-use note: each project targets a different niche — this compares them on the axes EyeBrowse optimizes for (stealth + LLM-drivable + one library/MCP codebase), not as an overall ranking.

Use as a library

import asyncio
from eyebrowse import EyeBrowse

async def main():
    eb = EyeBrowse()                           # stealth defaults
    try:
        async with eb.session() as s:          # a stealth session (auto-closed)
            await s.navigate("https://example.com")
            print(await s.snapshot())          # ARIA tree with [ref=...] handles
            await s.click("e6")                # act on a ref
            await s.type("e8", "hello", submit=True)
            png = await s.screenshot(full_page=True)
            title = await s.page.title()        # full Playwright power when you need it
    finally:
        await eb.aclose()

asyncio.run(main())

Run the included proof: uv run python examples/direct_usage.py.

Use over MCP

EyeBrowse ships an MCP server (eyebrowse-mcp, FastMCP over stdio). Add it to any MCP client.

Claude Code (CLI):

claude mcp add eyebrowse -- eyebrowse-mcp

Any MCP client (JSON config):

{
  "mcpServers": {
    "eyebrowse": {
      "command": "eyebrowse-mcp"
    }
  }
}

Then drive the loop: browser_navigate(url) → read the snapshot → act by ref (browser_click / browser_type / …). A default session is auto-created, so most tools just work. Full list: docs/TOOLS.md.

Proxy & identity (optional)

Runs proxyless by default (geoip still aligns locale/timezone to your real IP). Add a proxy only when you want one:

await eb.new_session(proxy="http://user:[email protected]:8080")
await eb.rotate_identity(proxy="socks5://host:1080")   # fresh fingerprint + paired IP
await eb.new_session(no_proxy=True)                     # force proxyless

Set a default once via EYEBROWSE_PROXY_* in .env, eb.set_static_proxy(...), or a custom ProxyProvider for rotation. Over MCP: browser_new_session(proxy_url=…) / browser_new_identity(proxy_url=…) / browser_set_proxy(…).

reCAPTCHA v3 / reputation gates are score-based and key off IP + session reputation — a fresh browser on a flagged IP fails regardless of stealth. Pair EyeBrowse with a clean residential proxy.

Extraction

eb.extract() (or browser_extract) hands the rendered HTML to Crawl4AI's raw: feed and returns clean, pruned markdownno LLM is called and no LLM keys are ever read; the consuming agent does any structuring.

md  = await eb.extract()                            # markdown string
res = await eb.extract(output_path="data/page.md")  # → {"path": ..., "chars": ...}

Recording

Native video — Playwright records the whole session to a .webm, written on close. The path is known up-front; the file finalizes when the session closes:

s = await eb.new_session(record_video=True)
# ... drive the browser ...
print(await s.video_path())          # path is known up-front; file finalizes on close
await eb.close_session(s.id)

Over MCP: browser_new_session(record_video=True)browser_video_path. Want a GIF for a README? Convert the .webm with ffmpeg (ffmpeg -i demo.webm demo.gif). The demo at the top was captured this way — see examples/make_demo.py.

How it works

CONSUMERS                         ENGINE (library: eyebrowse/)
 Claude Code  ──MCP──▶  mcp/  ──▶  EyeBrowse façade (public API)
 your code   ─ import ──────────▶   ├─ BrowserEngine (CloakBrowser / stealth Chromium)
 any MCP client                     ├─ proxy / identity rotation (pluggable)
                                    ├─ captcha solvers (pluggable, API-mode)
                                    └─ Crawl4AI (raw: feed) → clean markdown

The façade (EyeBrowse + Session) is the product; the MCP adapter is a thin 1:1 wrapper over it. The high-level API is curated and LLM-friendly — not a reimplementation of all of Playwright — and the raw page / context / browser objects are always one attribute away. The launcher is the only engine-specific layer; everything else is plain Playwright.

Caveats

Worth knowing:

  • evaluate runs in the page's main world (page globals reachable). To override a page's widget globals and fire a site callback (e.g. for captcha), EyeBrowse injects a <script> so the code runs in the page world — see captcha/inject.py.
  • HAR export closes the session — Playwright only flushes the HAR buffer when the context closes. Use the checkpoint pattern: browser_storage_statebrowser_har_exportbrowser_new_session(storage_state=...). For the initiator-rich Chrome HAR (JS call stacks), reach the Network.* domain via browser_cdp_send.
  • Native video is .webm — convert to GIF/MP4 with ffmpeg if you need another format.

Project layout

eyebrowse/
  api.py            EyeBrowse façade — the single public entry point
  config.py         settings / secrets (pydantic-settings)
  snapshot.py       aria_snapshot(mode="ai") + aria-ref= resolution
  proxy.py          ProxyConfig + pluggable ProxyProvider
  identity.py       Identity + random_identity() (isolated profile dir)
  extract.py        Crawl4AI raw: feed → markdown (lazy, optional dep)
  engine/           engine.py (CloakBrowser launch) + session.py (verbs + registry)
  captcha/          solver ABC + 4 providers + DOM detect/inject
  mcp/              FastMCP server + state + tools/ (18 groups · 85 tools)
examples/direct_usage.py   library proof (no MCP)
examples/make_demo.py      the native-video demo above
docs/TOOLS.md              full tool reference

Build notes, version-pin rationale, and verified engine behavior live in CLAUDE.md.

Use responsibly

EyeBrowse drives a real browser with anti-detection features. Use it only against sites you own or are explicitly authorized to automate, and within their terms and applicable law.

License

MIT © Evil-Bane


Found EyeBrowse useful? ⭐ Star the repo — it genuinely helps.

Built with Python · Playwright · CloakBrowser · FastMCP · the Model Context Protocol