NanoAgents lets you build and run small AI agents entirely in-browser using WebGPU — no API keys, no server calls, zero cost per query.
Stack: Next.js + WebLLM (Qwen2.5 0.5B–3B quantized models). Features include a skill chain system, real-time streaming output, short/long response modes, and a fork-and-customize workflow.
Everything runs on your GPU. The models are ~300MB–1.5GB and cached after first download. Works best in Chrome with WebGPU support.