Tilion — a stealth browser for AI agents

mcp-name: io.github.tiliondev/fortress

pip install tilion — drive a real, undetected Chromium that gets past Cloudflare, DataDome, and bot detection. No server, no account, no API key.

Framework & MCP: Beta · runs local & free · Tilion Cloud (residential egress) coming soon

Real, dated run against stockx.com (PerimeterX). A stock browser gets HTTP 403 — "Access denied"; an agent with the Tilion MCP returns clean JSON — same site, same prompt.

Tilion runs a recompiled-Chromium stealth engine (Fortress) locally, in-process, and gives you one clean API for fetching protected pages, extracting content, crawling sites, reconnaissance, and multi-step automation — plus a Model Context Protocol (MCP) server so AI agents can reach for it the moment they get blocked.

pip install tilion

Quickstart

import asyncio
from tilion import Tilion

async def main():
    async with Tilion() as t:                                 # boots Fortress locally
        page = await t.fetch("https://protected.example")     # past Cloudflare/DataDome/403
        print(page["title"], page["text"][:200])

        data = await t.extract("https://site/pricing")        # clean markdown + tables
        docs = await t.crawl("https://site", depth=2)         # whole-site crawl (auto-SPA)
        apis = await t.recon("https://site")                  # discover the site's private API
        hits = await t.search("undetected playwright")        # real-browser web search

asyncio.run(main())

No uvicorn, no Redis, no auth — local mode holds one real browser and drives it directly.

Drop-in stealth for your existing stack

Already on browser-use, Playwright, Puppeteer, or Crawl4AI? They connect to a browser by CDP URL. Point that at Tilion and your code runs through the stealth engine — one line, no rewrite:

t = await Tilion().start()
cdp_url = t.cdp_url                       # hand this to any CDP-speaking stack

from browser_use import Agent, BrowserSession
agent = Agent(task="...", browser_session=BrowserSession(cdp_url=cdp_url))

MCP server — stealth browsing for any AI agent

pip install "tilion[mcp]"
tilion-mcp                                # or:  npx -y tilion-mcp

Claude Desktop / Cursor / Cline / Windsurf — add to the MCP config:

{ "mcpServers": { "fortress": { "command": "tilion-mcp" } } }

Claude Code (CLI): claude mcp add fortress -- tilion-mcp

26 tools: fetch_protected_page, extract_page, crawl_site, recon_site_apis, search_web, run_browser_task, save_page, save_profile, get_stealth_cdp_endpoint, and more — pre-warmed (~100 ms first call), concurrency-safe, timeout- and SSRF-guarded. Full tool table →

What you get

How stealth works

The engine is a recompiled Chromium C++ fork (shipped as the tilion-fortress dependency), not a JS patch or a stealth plugin. Persona, User-Agent, WebGL, and canvas fingerprints are applied natively with genuine binding returns (toString() === [native code]), so there are no JavaScript injection tells. In independent suites it runs Sannysoft-clean, CreepJS 0% headless, BrowserScan "Normal."

Honest note: a great fingerprint on a datacenter IP still gets blocked by the biggest sites — real anti-bot decisions weigh the egress IP heavily. For hostile targets, add a residential proxy or use hosted Tilion cloud egress (coming soon). Tilion does not claim to be "undetectable."

Install options

pip install tilion              # core: stealth browser + fetch/extract/crawl/recon/search
pip install "tilion[mcp]"       # + the MCP server
pip install "tilion[vision]"    # + LLM-driven agent (Claude/OpenAI/Gemini)

Links

• Engine + MCP repo: https://github.com/tiliondev/fortress
• MCP setup & tools: https://github.com/tiliondev/fortress/tree/main/mcp
• Engine package: https://pypi.org/project/tilion-fortress/

Open-core: the tilion facade + MCP are BSD-3; the engine (tilion.core driver + the Fortress binary) is proprietary. Requires Python 3.10–3.13 on Linux, macOS, or Windows.