We all built that demo: a chatbot that answers questions about our data. Impressive for five minutes, useless on Monday morning.

So what separates a chatbot from an agent? The ability to act. And the Model Context Protocol is what makes it happen.

In this session we build live an agent that does real work. It reads from storage, calls a service, produces an artifact and publishes it. We'll write an MCP server from scratch, connect it to Gemini and watch the agent orchestrate tools it has never seen before.

But the happy path is the boring part! We'll explore what rarely reaches the stage: how to scope tool permissions so your agent can't do something expensive or irreversible, how to handle a tool that fails halfway through, how to stop the model inventing calls that don't exist, and how to deploy everything serverless on Cloud Run without paying for idle time.

You'll leave this session able to turn an internal API you already own into something an AI agent can safely use.

Your chatbot answers questions. Your agent should do the job! Let's build an MCP server live, plug it into Gemini and watch it act on real tools. Permissions that keep it out of trouble, failures handled properly, serverless deploy on Cloud Run. Come and turn an API you already own into something an agent can safely use!