Off-grid LLM over Radio

Messages flow both ways between the Meshtastic network and the model. Anyone on the mesh can send a prompt and get a reply, and the platform keeps message history so follow-up questions stay in context. It works much like a chat bot, just carried over radio instead of the internet. No signal? No problem.

A /tool prefix lets the model carry out actions rather than just answer questions. The project ships with one working example, a demo for calling emergency services, and the README describes how to add your own tool by editing the handler, registry, and config files. This is where the platform goes beyond a simple chat relay, and you'll do the wiring yourself.

Radio messages are short, so long LLM answers get split into chunks automatically. The platform handles the break at roughly 200 characters, which keeps replies readable across the mesh. You don't trim output by hand.

The bot can pull node-specific details such as battery level and location into its answers. For field teams, that means an AI reply can carry context about the device that asked the question. Handy when you're tracking several radios at once.

You can answer the whole channel or direct a response to one node, and that flexibility matters on a shared mesh, where a broadcast reaches everyone in range but a targeted reply keeps the airwaves from getting cluttered.

The platform supports USB, TCP, and Bluetooth links to your Meshtastic hardware. You set the interface in a few lines of code, so the same setup works whether your node sits next to the computer or hangs somewhere else in your network. Three connection types, one script.

It runs models through Ollama, so nothing leaves your machine and no cloud account is required. You pick the model by editing the chat function, and the README suggests llama 3.2:3b as a starting point. Pair it with a small model and even modest hardware can keep up. Bigger models mean better answers, if your machine can handle them.