LocalRouterGitHub

Your GPU. Your models. One router.

Fast & opinionated
local AI routing

One lightweight daemon. A single Zig binary. Generative models on your own GPU box, starting with NVIDIA DGX Spark.

View on GitHub

How it works

The right model. When you need it.

  1. 01

    One connection

    One MCP server for every agent. An OpenAI-style HTTP API for everything else.

  2. 02

    Room to run

    Models load on demand and unload when idle. Idle models make room by priority.

  3. 03

    Keep favourites warm

    Priorities keep your favourite models ready for the next job.

  4. 04

    Every model type

    Image and video today. Voice, music and more planned as adapters.

Local, from the kernel up

Built to generate.

Hand-written Zig engines with custom GPU kernels.

Qwen-Image 2.1

Text to image · 1024 × 1024 in about 12–17 seconds.

MiniMax H3

Text to video · 5 seconds at 768 × 448, with stereo audio.

Measured on a DGX Spark. Image editing and image to video are in progress.

Quick start

Install. Connect. Generate.

One command to install. One connection for your agent.

Terminal
git clone https://github.com/jayleaton/localrouter && bash localrouter/tools/spark/install.sh
claude mcp add --transport http localrouter http://<host>:8190/mcp

Replace <host> with your GPU box’s hostname or IP address.