fortress

MCP serverWeb & browsing

Stealth browser for AI agents: fetch pages behind Cloudflare/DataDome/CAPTCHA, extract clean data.

Unavailable. This server has no hosted endpoint yet, so ahel can't serve it.

Connect ahel once, and every AI you use reads what you have installed.

From the project's README

As published by tiliondev/fortress in README.md.

One browser engine to rule them all

Stealth Chromium engine

Fortress is a stealth Chromium engine that stops your scrapers and browser agents from getting blocked, with one line of code change. Bot detectors flag automation by reading the browser fingerprint; Fortress corrects that fingerprint inside Chromium's C++, so the browser presents as an ordinary Chrome install. Scrapers finish their runs, agents reach the pages they were sent to, and CreepJS, Sannysoft, BrowserScan, and live Cloudflare Turnstile all read it as human. Point your existing Playwright or Puppeteer at Fortress over CDP, and nothing else in your code changes.

Blink · V8 · BoringSSL patched in-tree · ANGLE / D3D11-backed WebGL · JA3/JA4-coherent TLS · monthly upstream rebase · reproducible, gauntlet-gated releases

Unedited capture of the Fortress binary in a real window: it clears a live Cloudflare challenge, turns bot.sannysoft.com all green, then reads BrowserScan “Normal”. Reproduce with tools/gauntlet.py.

Native-code parity

Every spoofed getter is a C++ getter: toString returns [native code], realm-invariant across main frame, iframes, and Web Workers.

Drop-in CDP

nodriver-style raw CDP on :9222, with no Runtime.enable leak. Keep Playwright, Puppeteer, or any CDP client; swap the browser, keep your code.

Clears the gauntlet

0% headless on CreepJS; Sannysoft, BrowserScan, and live Cloudflare Turnstile cleared, all as a stock Chrome install.

Auditable patches

34 small single-purpose diffs in patches/. Read one in a minute; rebuild the engine with one script.

Coherent by construction

Real V8, Blink, and BoringSSL keep engine, user-agent, and JA3/JA4 TLS shape in agreement: a Windows persona on a matching stack.

Tunable persona

One binary, a coherence-checked Windows identity; --uxr-* switches override any surface: GPU, screen, timezone, hardware, Client-Hints.


What's new — 151.0.7908.0 · Engine Refresh

Released. Per-launch coherent personas, hardened.

  • Per-locale keyboard (QWERTY / QWERTZ / AZERTY) · per-persona media devices · a full delivery-parity coherence pass.
  • Platform ↔ GPU ↔ timezone ↔ language ↔ voices ↔ keyboard move as one coherent real device — on every launch, from a single binary.
  • Native-code parity (toString() stays [native code]), realm-invariant (main / worker / iframe) — coherence compiled into the browser, not patched in JavaScript.
pip install -U tilion-fortress       # or:  docker run --rm -p 9222:9222 tilion/fortress:latest

→ full release notes

Contents

What it is · Quick startwhat it is, install, first script, AI-agent setup
The Fortress MCP29 stealth-browser tools for AI agents (Beta)
Why patch the engine, not the pagethe self-revealing-JS thesis + the three detection layers
How Fortress comparesvs puppeteer-stealth · Camoufox · CloakBrowser · closed vendors
Proof: live-detector resultsCreepJS / Sannysoft / BrowserScan / Cloudflare, with screenshots
Configure the personathe --uxr-* fingerprint surface
Works with your stackbrowser-use · Crawl4AI · Stagehand · LangChain
Build & verifyreproduce from source, verify provenance
Referencetroubleshooting · FAQ · roadmap · repo layout

What it is

Fortress is a Chromium fork that spoofs the browser fingerprint from inside the engine. The surfaces bot detectors read (canvas, WebGL, audio, fonts, navigator, and about thirty more) are corrected in Chromium's C++, with no JavaScript patch layer sitting on top for a page to catch.

It ships as an ordinary browser binary that exposes a CDP endpoint. Point Playwright, Puppeteer, or any CDP client at it and your existing automation runs unchanged.

A JavaScript stealth patch leaves an extra layer the page can find: .toString() shows the override's source, and re-grabbing the same primitive from an iframe or worker reaches past it. Fortress corrects the surface in the engine instead, so navigator.vendor resolves to the real C++ getter, reports [native code], and reads the same from every realm. A page inspecting itself sees stock Chromium. That is why your automation gets through where it used to get flagged, and whatever blocking is left traces to your proxies and behavior rather than the browser. Why patch the engine, not the page covers the detection mechanics in full.

from tilion_fortress import Fortress
from playwright.sync_api import sync_playwright

with Fortress() as f:                                   # launches the stealth engine on a CDP endpoint
    with sync_playwright() as p:
        browser = p.chromium.connect_over_cdp(f.cdp_url)
        page = browser.new_page()
        page.goto("https://bot.sannysoft.com")
        page.screenshot(path="all-green.png")
import { Fortress } from "tilion-fortress";
import { chromium } from "playwright";

const f = await Fortress.launch();                      // stealth engine on a CDP endpoint
const browser = await chromium.connectOverCDP(f.cdpUrl);
const page = await browser.newPage();
await page.goto("https://browserscan.net");
await browser.close();
await f.close();

The 12-second tour

One loop, all real captures: passes CreepJS / Sannysoft / BrowserScan / rebrowser → scrapes structured data over CDP → clears Akamai on aa.com · lowes · macys · kohls (same residential IP).

Real scraping, fully headless

Unedited captures of the Fortress engine driven over CDP. No stealth plugins, no JS patches: the fingerprint is corrected in the binary. Reproduce any of these with examples/scrape_demos.py.

Structured extraction: records build into typed JSON as each item is read.

Clears real Akamai — before / after

Same residential IP, same site (aa.com · Akamai Bot Manager). A stock/headless browser gets Access Denied (Reference #); Fortress loads the real page and Akamai issues its _abck sensor cookie — the Bot Manager accepts it as a real browser. The variable is the fingerprint, not the IP.

Not a one-site fluke — same before/after on major Akamai-protected retailers, every run from the same residential IP.


Quick start

# Python / Node: prebuilt native binary auto-fetched (Linux x64 & Windows x64), SHA-256 verified
pip install tilion-fortress
npm  install tilion-fortress

# Any OS via Docker: raw CDP on :9222  (~302 MB pull / 851 MB on disk, stripped single-layer)
docker run --rm -p 9222:9222 tilion/fortress:latest

# Portable bundle (extract-and-run, like a Chromium snapshot)
tar xzf tilion-fortress-linux-x64.tar.gz            # Linux
./tilion-fortress/tilion --headless=new --remote-debugging-port=9222 --user-data-dir=/tmp/p

# Debian / Ubuntu
sudo apt install ./tilion-fortress_151.0.7908.0_amd64.deb && tilion https://example.com

[!TIP] The SDK ships the compiled build plus patches/, so you can rebuild the engine yourself and verify every surface correction against the source. Downloads are SHA-256-verified against the release SHA256SUMS automatically.

Versions

Fortress ships on two Chromium bases — pick your trade-off between blend-in and currency:

ChannelChromiumWhen to use
stable (default)149Recommended — matches the Chrome version the mass of real users run, so it blends in best
latest151Newest engine (reports a version slightly ahead of stable)
Fortress().start()                     # Python: stable (149) by default
Fortress(channel="latest").start()     # opt into 151
await Fortress.launch();               // Node: stable (149) by default
await Fortress.launch({ channel: "latest" });
docker run --rm -p 9222:9222 tilion/fortress:149   # or :151
# or set FORTRESS_CHANNEL=latest for either SDK

Native binaries: Linux x64 (both versions) + Windows x64 (151); Windows-149 and macOS run via the Docker image.

Drop it into your AI agent

Fortress is the browser your agent drives: raw CDP on :9222, no stealth plugins to wire up. There are two ways in.

Option 1: open it pre-loaded in a chat assistant. One click; it reads our AGENTS.md and walks you through the whole setup:

Option 2: Copy for agent (everything, to your clipboard). Hit the copy icon at the top-right of the box below. It puts the entire setup context on your clipboard: what it is, install, connect, persona, and rules, all of AGENTS.md condensed. Paste it into Cursor, Claude Code, Copilot, ChatGPT, or any agent and it takes it from there:

You're setting up Fortress, an open-source STEALTH Chromium engine, for browser automation.
It corrects the browser fingerprint (canvas, WebGL, audio, fonts, navigator, +30 more) in Chromium's
C++ and exposes raw CDP on http://localhost:9222, a drop-in for Playwright/Puppeteer. Do NOT add
puppeteer-stealth or any JS fingerprint patching (it self-reveals and undoes Fortress).

LAUNCH (pick one; all expose CDP on http://localhost:9222):
  Docker:  docker run -d --rm -p 9222:9222 tilion/fortress:latest
  Python:  pip install tilion-fortress    then  from tilion_fortress import Fortress; f=Fortress(); f.start()
  Node:    npm install tilion-fortress    then  import {Fortress} from "tilion-fortress"; const f=await Fortress.launch()

CONNECT (keep my existing automation code):
  Playwright(py):  browser = p.chromium.connect_over_cdp("http://localhost:9222")
  Playwright(js):  const browser = await chromium.connectOverCDP("http://localhost:9222")
  Puppeteer(js):   const browser = await puppeteer.connect({ browserURL: "http://localhost:9222" })
  browser-use / Crawl4AI / Stagehand / LangChain:  point their CDP endpoint at http://localhost:9222

PERSONA (optional; default is a coherent Windows identity). Override any surface with --uxr-* flags:
  --uxr-timezone=America/New_York --uxr-hw-concurrency=16 --uxr-languages=en-US,en

RULES:
  1) Drive over raw CDP (:9222); don't spawn chromedriver.
  2) Never pass --user-agent (use --uxr-ua-*); it desyncs UA vs UA-Client-Hints.
  3) No puppeteer-stealth / undetected-chromedriver / JS fingerprint patches.
  4) Blocked ~90% = my IP (datacenter), not the fingerprint. Use a residential/mobile proxy, then retry.

Now walk me through launching Fortress and wiring my automation to it.
Full guide: https://github.com/tiliondev/fortress/blob/main/AGENTS.md

The Fortress MCP — stealth browsing as agent tools  Beta

Raw CDP is for code you write. The Fortress MCP is for agents that call tools: a Model Context Protocol server that hands Claude, Cursor, or any MCP client a stealth browser, so the moment a fetch is blocked it just calls a tool and gets the page. 29 tools, local and freefetch_protected_page, extract_page, crawl_site, recon_site_apis, search_web, run_browser_task, save_profile, get_stealth_cdp_endpoint, and more.

Real, dated run against stockx.com (PerimeterX). A stock browser gets HTTP 403 — “Access denied”; an agent with the Fortress MCP returns clean JSON — same site, same prompt. Reproduce it from the framework repo.

Set it up in 30 seconds

Two runners — pick one. npx needs Python on PATH; pip installs it directly:

pip install "tilion[mcp]"      # command:  tilion-mcp
#   —or, zero-install—
npx -y tilion-mcp              # auto-runs the server via uv (no global install)

Claude Desktop — Settings → Developer → Edit Config (claude_desktop_config.json):

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

Prefer npx? Use "command": "npx", "args": ["-y", "tilion-mcp"]. Restart Claude, and the fortress tools appear.

Claude Code (CLI) — one line:

claude mcp add fortress -- tilion-mcp          # or:  claude mcp add fortress -- npx -y tilion-mcp

Cursor (~/.cursor/mcp.json) · Cline / Windsurf (VS Code → MCP servers) — same block:

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

Then just ask your agent — “get the price off this StockX page” — and it calls fetch_protected_page on its own.

What the agent gets

tools
Get blocked pagesfetch_protected_page · read_page · get_page_html · search_web
Structured dataextract_page (schema-aware) · extract_document (PDF/DOCX/XLSX)
Whole sitescrawl_site (auto-SPA) · recon_site_apis (find the private JSON API)
Drive a pagepage_elements · click_button · fill_field · press_key · wait_for · evaluate_js
Multi-step flowsrun_browser_task (login, paginate, infinite-scroll, checkout, …)
Capture / authscreenshot_page · save_page · download_file · save_profile / load_profile
Bring your ownget_stealth_cdp_endpoint → a CDP url for Playwright / Puppeteer / browser-use

Tools are annotated (reads auto-approve, writes gate), pre-warmed on startup (~100 ms first call), concurrency-safe, and timeout- and SSRF-guarded. A hosted endpoint with residential egress is coming soon.

→ Full 29-tool table, benchmarks, and the agent skill: mcp/


Why patch the engine, not the page

The usual approach patches navigator.webdriver, spoofs the WebGL vendor, and overrides navigator.plugins from script. CreepJS and similar detectors still flag it, and the reason is structural, not one more property left uncovered. A JavaScript spoof is a function standing where a native one belongs. Detectors set the returned value aside and interrogate whether the thing returning it is native:

The tellWhy it catches a JS spoof
toString self-revealA native method stringifies to function get vendor() { [native code] }; an override stringifies to its own source, so one .toString() catches it.
Descriptor and hasOwnPropertygetOwnPropertyDescriptor exposes redefined props, and hasOwnProperty('toString') returns true on a tampered function where a native one returns false.
failsTypeErrorNative getters throw a specific TypeError on the wrong this; a naive shim stays quiet, and the silence is the signal.

Realm re-acquisition is the one that defeats every main-world patch. A detector grabs a pristine primitive from another realm and turns it on your function:

const iframe = document.createElement('iframe'); document.body.appendChild(iframe);
const realToString = iframe.contentWindow.Function.prototype.toString;
realToString.call(navigator.__lookupGetter__('vendor')); // returns your source code. Caught.

Your main-world patch lives in a different realm from that iframe. The same trap fires from a Web Worker, a thread your main-thread shim runs beside rather than inside.

Fortress has no such layer. The getter for navigator.vendor is the C++ getter: it reports [native code] because it is native code, identical across every realm. Camoufox puts it well: "there is no JavaScript hijacking to be detected." Fortress applies the same idea to V8 and Blink in place of Gecko.

The three layers of bot detection, and where Fortress fits

Modern anti-bots (Cloudflare, DataDome, Kasada, HUMAN, Akamai) read three structurally different surfaces, in three separate places. One tool rarely fixes all three:

Shortened here. Read the whole README on GitHub.

Signals

GitHub stars
488
Forks
35
Last commit
Jul 2026
Advanced
Delivery
fortress MCP server → your ahel gateway (mcp.ahel.ai) → every connected AI client.
Catalog kind
mcp-server
Gateway key
io-github-tiliondev-fortress
Source
github.com/tiliondev/fortress