Subwave: an agentic DJ powered by Ollama and a 9B model runs fully local

Subwave transforms a local Ollama server into a fully autonomous radio DJ. The system uses an agentic loop: at each turn, the model calls tools to search the Navidrome music library, review recent plays, check the schedule, and fetch weather data from Open-Meteo. It then selects the next song with a written reason, generates a short intro, passes it to a local TTS engine (Piper or Kokoro), and uses Liquidsoap to duck the music volume under the voice. The creator found that a 9B parameter model (Qwen3.5 9B) with reasoning turned off and tool‑calling enabled is sufficient for the task. Larger models write better copy but are not required. The critical factor is session memory: without it, the DJ repeats itself within an hour. Subwave runs fully on your own hardware—Ollama for the LLM, Navidrome for the music library, and any Docker host. The only outbound call is to a free weather API. Hosted LLM providers are optional and capped by a daily token budget.

Subwave is released under the MIT license on GitHub, and a live demo is available at getsubwave.com. The creator notes that the experience is radio—there is no skip button, which some listeners love and others immediately reject. This project demonstrates how small, locally‑run models can power engaging, autonomous applications without cloud dependencies or per‑user costs. For professionals experimenting with local AI, Subwave offers a practical template for building agentic tools that interact with real‑world services and media libraries.