N0X
Private AI workbench for files, research, and local LLMs.
Start new thread

N0X - Run any LLM in your browser — offline, private, zero cloud.

by•
Since launch 1: added a ReAct agent loop that runs tools (web search, Python, docs, image gen) autonomously. Hybrid RAG now uses BM25 + vector search fused with RRF + MMR reranking way better recall than vector-only. Auto-routing picks local vs cloud per message complexity. Chrome AI (Gemini Nano, zero download) works alongside WebGPU, Ollama, and any OpenAI-compatible endpoint. Persistent semantic memory across sessions. GPU-tier detection now blocks models that'd OOM your device.

Add a comment

Replies

Best
Maker
📌
Built this because I was tired of pasting sensitive docs into ChatGPT and just hoping for the best. Started as a weekend experiment "how hard is it to run Llama in a browser tab?" Turns out, hard. WebGPU is wild. Half my time went into worker thread hell and figuring out why the model would just freeze at 0% forever. The part that surprised me most: hybrid search (BM25 + vector + reranking) on documents actually works really well in WASM. I expected it to be garbage. It's not. This is the 2nd launch because the first version was basically just "load a model, chat with it." Version 2 added an agent that can actually browse, run code, search your docs, and generate images all chained together, no backend. Happy to answer anything. Especially if something's broken on your GPU that's still the hardest part to debug remotely.

Love that it actually keeps everything local with browser-side inference, no sketchy data grabs or surprise paywalls. The Ollama and OpenAI-compatible fallback is a smart move for when you need a bigger model without rebuilding the whole workflow.

how does the browser-side inference actually hold up on integrated GPUs, like is it usable on a regular macbook air or are we talking proper discrete card territory?