Every prompt you send to a cloud coding assistant takes your source code, customer data and secrets with it. So I started building Java applications with a language model running on my own laptop, with no cloud account and no dedicated graphics card. This talk shares what I learned along the way, and shows how to do it well.

Starting from nothing, I'll set up llama.cpp, pick a model that fits the hardware, and connect a coding agent. Then we'll build a small Java application live, with the network unplugged to prove that nothing leaves the machine.

You'll see what makes the difference in practice: how to choose a model, how to work within a small context window, how to keep the agent on track with tests and small steps, and how to review what it produces.

I'll also be honest about the limits: local models are slower and weaker than the big cloud services, they make mistakes you need to catch, and an agent with tools or web access can still leak your data. You'll leave with a setup you can try yourself, realistic expectations and a clear idea of when local is and isn't the right choice.