⚡ ModelNorth HyperBrowser

The Sovereign Tri-Tier Web Agent Engine.
Sub-20ms reflex execution, $0.00 local base cost, and multimodal vision sentry fallbacks for Canvas and CAPTCHAs.

Live execution of the sovereign Dubai (DXB) → Lahore (LHE) benchmark running on Tier 1 Local Reflex with zero cloud LLM latency.

---

🚀 Key Advantages

• ⚡ Sub-20ms Reflex Loop: Eliminates the 5–15 second multimodal LLM latency bottleneck for standard web navigation.
• 💰 $0.00 Base Cost: Runs Tier 0 (in-V8 intent compiler) and Tier 1 (HyperLocal ONNX INT8) 100% locally on your machine with zero cloud API keys.
• 🧬 Recursive Shadow DOM & iframe Crawling: Seamlessly pierces web components, closed/open shadow roots, and embedded frames.
• 🕵️ Anti-Bot Stealth Layer: Integrated CDP evasion masking navigator.webdriver, injecting realistic Bezier curve mouse movements, and spoofing WebGL signatures.
• 📊 Zero-Token Structured Extraction: Instantly extracts DOM tables and cards into validated Pydantic schemas in $< 30\text{ ms}$.
• 👁️ Multimodal Vision Sentry: Seamlessly falls back to Gemini 2.5 Flash for <canvas>, WebGL, interactive charts, and visual CAPTCHAs with exact pixel coordinate grounding.
• 🛡️ 100% Sovereign & Private: No DOM structures, form values, customer credentials, or session cookies are sent to third-party decision APIs.

---

🏛️ Tri-Tier Architecture

                       USER GOAL
                           │
                           ▼
   ┌───────────────────────────────────────────────┐
   │         Browser Session Harness (CDP)         │
   │  • Anti-bot stealth & Bezier mouse dynamics   │
   └───────────────────────┬───────────────────────┘
                           │ Page State & Geometry
                           ▼
   ┌───────────────────────────────────────────────┐
   │ TIER 0: In-Browser Heuristic Compiler (V8)    │ ➔ Latency: < 1 ms  |  Cost: $0.00
   │ Direct intent matching & Shadow DOM crawl     │
   └───────────────────────┬───────────────────────┘
                           │ If Ambiguous
                           ▼
   ┌───────────────────────────────────────────────┐
   │ TIER 1: Local System 1 Decision (HyperLocal)  │ ➔ Latency: 7–15 ms |  Cost: $0.00
   │ Speculative action classification & ranking   │
   └───────────────────────┬───────────────────────┘
                           │ If Canvas / CAPTCHA / Stuck
                           ▼
   ┌───────────────────────────────────────────────┐
   │ TIER 2: Sovereign Vision Sentry (Gemini Flash)│ ➔ Latency: ~300 ms |  Cost: ~$0.00005
   │ Exact [ymin, xmin, ymax, xmax] pixel clicks   │
   └───────────────────────────────────────────────┘

---

📦 Quickstart

1. Installation

git clone https://github.com/ModelNorth/modelnorth-hyperbrowser.git
cd modelnorth-hyperbrowser
uv sync

2. Run Preflight Health Suite

uv run python scripts/preflight_check.py

3. Run the Karpathy AutoResearch Loop

uv run python scripts/autoresearch_runner.py --iterations 3

---

💻 Python Library Usage

Fast Action Execution

from modelnorth import HyperAgent

goal = (
    "Find one-way flights from Zurich to London on September 20, 2026, "
    "for one adult in economy. Stop when flight list is visible."
)

async with HyperAgent(url="https://www.google.com/travel/flights?hl=en", goal=goal) as agent:
    async for step in agent.run():
        print(f"[Tier {step.tier}] {step.action} -> {step.target_name} ({step.elapsed_ms:.1f}ms)")

Zero-Token Structured Data Extraction

from pydantic import BaseModel
from typing import Optional
from modelnorth import HyperAgent

class FlightCard(BaseModel):
    airline: Optional[str] = None
    price: Optional[str] = None
    stops: Optional[str] = None

async with HyperAgent(url="https://www.google.com/travel/flights?hl=en", goal="Show flights") as agent:
    # Extract records directly without LLM token cost (< 30ms)
    flights = await agent.extract(schema=FlightCard)
    for flight in flights:
        print(flight)

---

📊 Benchmark Comparison

Metric	Traditional Vision Agent	Legacy Cloud Agent	ModelNorth HyperBrowser
Standard Step Latency	5,000 – 15,000 ms	200 – 500 ms	1 – 15 ms (Local)
Cost per 1,000 Steps	$50.00 – $250.00	$5.00 – $15.00	$0.00 – $0.05
Canvas / WebGL Support	✅ Yes	❌ Fails	✅ Yes (Tier 2 Sentry)
Visual CAPTCHAs	✅ Yes	❌ Fails	✅ Yes (Tier 2 Sentry)
Local Offline Mode	❌ No	❌ No	✅ Yes (100% Offline)

---

📜 License

Licensed under the Apache License, Version 2.0. See LICENSE for details.
Copyright © 2026 ModelNorth Sovereign AI.