# warpdrv

An alpha-stage desktop toolkit that manages local llama.cpp server instances with a rich chat UI, voice, retrieval-augmented generation, sub-agents and an OpenAI-compatible proxy.

warpdrv is a desktop toolkit for using local large language models. It launches and manages llama-server instances with full parameter control, including cache checkpointing, speculative decoding, n-gram speculation and multimodal projections, with per-model settings saved so launching a model is a single click. Multiple llama-server builds, such as stock, custom-compiled, ROCm or CUDA variants, can be registered and grouped for quick swapping, and a recipe engine can compile fresh builds on demand.

It provides an OpenAI-compatible proxy endpoint so any chat app can talk to a running server through named aliases, with optional authentication and a built-in web server for remote access. Its own chat UI supports elicitation, graphs and syntax highlighting, and it adds voice through text-to-speech and dictation, retrieval-augmented generation over a workspace with local embeddings, and sub-agents that delegate tasks in parallel across CUDA, ROCm or Vulkan hardware.

It is an alpha release under active development, written in TypeScript.

Website: https://github.com/mikjee/warpdrv