idosal/AgentLLM

Leverages WebGPU for GPU-accelerated inference of open-source LLMs (WizardLM) directly in the browser, eliminating server dependencies. Implements an autonomous agent loop inspired by AgentGPT that generates and executes tasks without external tools, simplifying reasoning while validating embedded model capabilities. Built on WebLLM and compatible with Chromium M113+, enabling privacy-preserving autonomous task decomposition entirely client-side.

457 stars. No commits in the last 6 months.