eyebrowse 0.1.6

EyeBrowse — a stealthy, LLM-drivable browser-control engine (Camoufox), consumed as a library and over MCP.

🦊 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 — Camoufox, a Firefox fork with C++-level fingerprint spoofing — so legitimate automation isn't false-flagged or IP-banned by Cloudflare, DataDome, Akamai, or PerimeterX.

_{▶ Higher-quality MP4: docs/demo.mp4 — real EyeBrowse captures, composed into an ad by examples/make_demo.py}

---

Why EyeBrowse?

• 🥷 Stealth by default — engine-level fingerprint spoofing (geoip + humanize + block_webrtc on out of the box); navigator.webdriver masked; viewport auto-sized to the spoofed screen. No puppeteer-extra band-aids — the anti-detection is in 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.
• 🧰 Library and MCP from one codebase — a clean Python API (EyeBrowse + Session), mirrored 1:1 by a thin MCP server (78 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, full HAR capture, screen recording (MP4/GIF), 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 · Library · MCP · Proxy & identity · Extraction · Recording · How it works · Caveats · Tools · License

Quickstart

pip install eyebrowse
python -m camoufox fetch          # one-time: download the stealth Firefox + GeoIP db

import asyncio
from eyebrowse import EyeBrowse

async def main():
    eb = EyeBrowse()                          # stealth defaults: geoip · humanize · block_webrtc
    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
python -m camoufox fetch             # one-time browser fetch (like Playwright's `playwright install`)
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)
uv run python -m camoufox fetch
cp .env.example .env                 # only if you use a proxy / captcha keys

Python 3.12 (pinned <3.13). Pinned engine: camoufox==0.4.11 (Firefox v135), playwright 1.60, mcp 1.27.

Features

🥷 Stealth	Camoufox engine-level fingerprint spoofing; geoip + humanize + block_webrtc by default; webdriver masked; viewport matched to the spoofed screen.
🤖 LLM interaction	aria_snapshot(mode="ai") → ARIA tree + [ref] handles; click / type / hover / select / drag / file-upload / dialogs / keyboard; coordinate mouse too.
🪟 Frames & DOM	cross-origin iframe routing by ref, shadow-DOM piercing, popup/new-tab switching, evaluate inside any frame.
🗂 Multi-session	independent stealth sessions, each with its own context / identity / proxy.
🌐 Network	inspect requests/responses (incl. XHR/fetch bodies & WebSocket frames), block URLs, mock responses, go offline, full HAR export.
💾 State	cookies, localStorage & sessionStorage (CRUD), storage_state save/reload.
🪪 Identity rotation	fresh fingerprint (OS/screen) + isolated profile + paired proxy; pluggable residential ProxyProvider.
🧩 Captcha	pluggable API-mode solvers (CapSolver / 2Captcha / CapMonster / NextCaptcha) + TOTP — no browser extension.
📄 Extraction	Crawl4AI raw: feed → clean, token-efficient markdown (no LLM, no API keys).
🎥 Capture	screenshots, Playwright tracing, and screen recording → MP4 / WebM / GIF (ffmpeg).
✅ Verify & debug	assertions, element highlighting, locator generation, geolocation/header emulation.

Full per-tool reference: docs/TOOLS.md (78 tools across 17 groups).

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:[e-mail]: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(…).

Extraction

eb.extract() (or browser_extract) hands the rendered HTML to Crawl4AI's raw: feed and returns clean, pruned markdown — no 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

Camoufox can't write native browser video, so EyeBrowse rolls its own recorder: it captures viewport frames and encodes them with ffmpeg into a smooth, real-time MP4/WebM (and a palette-optimized GIF that autoplays inline on GitHub). ffmpeg ships bundled (imageio-ffmpeg), so it works out of the box.

await s.start_recording(fps=30)
# ... drive the browser ...
await s.stop_recording("demo.mp4", extra_paths=["demo.gif"])

The demo at the top was produced 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 ──────────▶   ├─ Camoufox engine (stealth context, HAR)
 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.

Camoufox caveats

These shape the API — worth knowing:

• evaluate runs in an isolated world — it sees the DOM but not page-script window.* globals. Read app state via DOM / network / HAR.
• reload / back / forward aren't reported by Firefox's Juggler, so they're driven via goto + an internal navigate() history (they traverse navigate()-driven history, not click-driven).
• Not available on Firefox/Camoufox (intentionally omitted, not broken): page.pdf(), media emulation, and native Playwright video — use the ffmpeg recorder instead.

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()
  extract.py        Crawl4AI raw: feed → markdown (lazy, optional dep)
  engine/           camoufox_engine.py (launch) + session.py (verbs + registry)
  captcha/          solver ABC + 4 providers + DOM detect/inject
  mcp/              FastMCP server + state + tools/ (17 groups, 78 tools)
examples/direct_usage.py   library proof (no MCP)
examples/make_demo.py      the screen-recording demo above
docs/TOOLS.md              full tool reference

Build notes, version-pin rationale, and verified Camoufox 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