n0x / The Browser-Native AI Stack

Zero backend. 100% private. Blistering fast local inference.

Live Demo · Architecture · Deployment

n0x is a demonstration of pushing edge compute to its absolute limits. Utilizing WebGPU, WASM, and IndexedDB, n0x runs a complete AI ecosystem entirely within the client's browser. No API keys, no cloud compute costs, and zero data leaves the user's local machine.

Technical Capabilities

• WebGPU Inference Engine: Direct-to-metal LLM execution via MLC/WebLLM. Hits 40+ tokens/sec on standard consumer hardware. Models (Llama 3, Qwen 2.5, Phi 3.5) are downloaded once and aggressively cached.
• WASM Vector Search (RAG): Drag-and-drop document processing. Chunking, all-MiniLM-L6-v2 embedding generation via transformers.js, and cosine similarity search via voy-search — all running locally in WebAssembly.
• Sandboxed Python Runtime: Client-side execution of Python data analysis and scripting via Pyodide. Code output flows directly back into the LLM context.
• Persistent Telemetry & Memory: Long-term conversational memory and engine orchestration state persisted securely via IndexedDB.

Architecture & DAG Execution

n0x uses an event-driven architecture to orchestrate models and APIs.

[User Input] → [Intent Router]
                   │
                   ├─→ [Deep Search Enabled?] ──→ Serverless Proxy (DuckDuckGo/Tavily)
                   │
                   ├─→ [Documents Attached?] ──→ WASM Vector DB (Voy) → Extract Context
                   │
                   ├─→ [Python Requested?] ────→ Pyodide (WASM Sandbox) → Output Capture
                   │
                   └─→ [Memory Context] ───────→ IndexedDB Vector Retrieval
                             │
                             ▼
              [WebGPU Base LLM (e.g. Qwen 2.5)]
                             │
                             ▼
                   [UI Render / Telemetry]

Quick Start (Development)

Requires Node 18+ and Chromium 113+ (for WebGPU API support).

git clone https://github.com/ixchio/n0x.git
cd n0x
npm install
npm run dev

Open localhost:3000. The engine will initialize and download the default quantized weights on the first load.

Models & Memory Footprint

Weights are quantized (q4f16) for optimal VRAM usage.

Tier	Default Weights	VRAM Est.	Tokens/Sec (M-series)
Fast	Qwen2.5 0.5B Instruct	~400MB	50+
Balanced	Qwen2.5 1.5B Instruct	~1.1GB	35+
Powerful	Llama 3.2 3B Instruct	~2.2GB	20+
Code	Qwen2.5 Coder 1.5B	~1.1GB	35+

Security & Privacy

This project is built on a "local maximum" privacy thesis.

Because the inference graph and orchestration layer remain entirely in the browser, PII and proprietary code never transit a network boundary. The only exceptions are explicit external hooks:

• Search Retrieval: Pings a Next.js serverless route proxy to bypass CORS.
• Image Generation: Pings the free Stable Horde / Pollinations API. Both can be toggled off individually, guaranteeing a 100% air-gapped run state.

Stack

Next.js 14 TypeScript WebLLM (WebGPU) Pyodide (WASM) Transformers.js Tailwind CSS Framer Motion Zustand

License

MIT © ixchio