An AI assistant that can work the bands with you.

Why

One of the most interesting things AI can do is operate autonomously in fuzzy environments. We have self-driving cars in messy city traffic, drones that find their own targets, lawn mowers that know where the lawn ends and the flower bed begins.

Radio communication is fuzzy by its very nature. Humans talk outside of fixed schemas, and antennas, power, band conditions and QRM all alter what the receiver sees. Wireless communication has used error correction for decades, but that works by transmitting redundant information and using math to reconstruct the message, or to request it again. The challenge here is different: have an AI carry a QSO the way a human does, preferably in a way that appears human on the receiving end.

Skeptics will say Part 97 requires human operators. Not quite: it requires operator supervision, and machine communication has plenty of precedent, from Wi-Fi to FT8 to self-announcing beacons. Does autonomous operation not defeat the point of a hobby built on human operators? Anyone who has been in this hobby for decades knows it is mostly about the tinkering: once the new rig is going and the antenna is perfect, we move on to the next shiny thing. This project is not about replacing human operators. It is about experimenting and demonstrating what is possible.

I started this project in 2026, preparing for Field Day, as a simple proof of concept. It worked: we operated the very first autonomous QSO in history, with me as the licensed operator at the radio, ready to take over at any moment.

Agentic Connectors

At its core are three Model Context Protocol (MCP) connectors. MCP servers are for AI what drivers are for hardware: they let an AI talk to the real world, to software, or even to a radio.

With these tools and some agentic plumbing, an AI can initiate or answer a QSO and log it automatically, all day, while you supervise and enjoy whiskey and cigars.

What is next

During Field Day, the agentic plumbing was a frontier model, cloud-hosted, which requires an always-on Internet connection. That has three disadvantages: you have to be connected to the Internet to communicate via a radio, which rather defeats the purpose; you are eating up tokens; and frontier models are powerful but slow.

Instead of running a trillion-parameter model, the goal is a small billion-parameter model that can run in the field, on a modern CPU with AI extensions or on an inference processor like those some drones carry. What we need is a purpose-built ham-radio model. Its sole purpose is to understand how humans communicate on the digital modes and to do smart error correction, very much the way our human mind reads the right message out of a partially garbled decode. The first step in building a model from scratch is collecting training data. More on the Ham Radio Model page.