A fully local agentic AI radio DJ running on a 9B model
A developer has built and open-sourced "Subwave", a fully local agentic radio DJ that uses Ollama, Navidrome, Piper/Kokoro TTS, and Liquidsoap. The system uses a 9B model (such as Qwen) with tool-calling to search music libraries, check the weather, and write voice intros, while Liquidsoap handles the audio mixing. This project demonstrates that smaller, local 9B models are highly capable of handling complex agentic workflows when equipped with tool-calling and session memory. It offers a privacy-focused, API-key-free alternative to cloud-based AI audio generation for self-hosted media enthusiasts. The system relies heavily on session memory to prevent the DJ from repeating tracks within a short timeframe, proving more critical than model size. It runs entirely on local hardware using Docker, requiring only a single outbound API call to Open-Meteo for weather updates.
## BACKGROUND
Navidrome is an open-source, self-hosted music server, while Liquidsoap is a specialized scripting language designed for generating and automating audio streams. Kokoro TTS is a lightweight, open-source text-to-speech model, and Ollama is a tool that simplifies running large language models locally.