This chapter brings together tools, agents, and the Model Context Protocol (MCP) — the standard for exposing tools to LLMs as reusable services.
We build a KU Parking Assistant that:
Uses real deterministic tools (Haversine distance, fuzzy building matching, Google Maps URLs)
The LLM decides when to call which tool based on the question
The same tools are deployed to a Google Cloud Run container as an MCP server
Both a Python notebook and a Dify workflow can call the same tools
Two notebooks contrast:
Inline tools — tools defined as Python functions in the notebook
MCP tools — tools accessed over HTTPS from the Cloud Run endpoint