The Bulka chat on a local model

The Bulka chat agent can run on a local model: LM Studio, Ollama or any other OpenAI-compatible server on your computer. No cloud, no keys, with the full tool set (the model writes code and plays music itself). Works in the desktop app.

Quick start

1. Start the local model server:
   • LM Studio → Developer tab → Local Server → Start. Address: http://localhost:1234/v1.
   • Ollama → starts by itself. Address: http://localhost:11434/v1.
2. In Bulka: Chat tab → settings → the “Local (LM Studio / Ollama)” provider.
3. Enter the server address (there are LM Studio / Ollama preset buttons), press “Load models” and pick a model.
4. Talk: the model writes code and drives the music right in the editor.

How it works

Requests go through the built-in Bulka server to your local server (server to server), so CORS does not get in the way (a browser talking to LM Studio/Ollama directly is blocked by CORS). A key is usually not needed, but there is a field for it.

Recommendations

• To drive Bulka, the model needs function calling (tools) support: for example Qwen 3 / Qwen 2.5, Llama 3.x, Mistral. Tested on Qwen 3.8 27B: it sets the code and starts the track itself.
• Reasoning models show their thoughts in the chat as a collapsible block.
• The bigger the model, the more meaningful the music, and the higher the hardware requirements.