Member-only story
Building a local AI agent that uses Google Maps and Open-Meteo through tool calling
When Google introduced the Agent Development Kit (ADK), most of the examples used Gemini models hosted on Google’s infrastructure. The framework itself, however, is model-agnostic, and one of the first questions that came to mind was whether the same programming model could be used with a completely local language model.
The release of Gemma 4 through Ollama provided the perfect opportunity to try exactly that.
My goal wasn’t to compare models or benchmark inference speed. I wanted to understand how ADK approaches tool calling and whether a locally hosted model could orchestrate multiple external services in a reliable way. In other words, could Gemma reason about a task, decide which tool to invoke, pass the output of one tool into another, and produce a structured response without any custom orchestration code?
To answer that question I built a small weather agent. Weather may sound like an overly simple example, but it has one important characteristic: the model cannot possibly know the correct answer. Unlike questions about programming languages or history, today’s weather isn’t part of the model’s training data. The only way to answer correctly is to…