From API to Agent Tool: Connecting Your Apps with MCP
Your apps already have APIs — MCP is how AI agents get to use them safely. This talk shows how to wrap an existing API into an MCP server with Python, scope what agents can read and write, and connect coding agents locally or remotely, demonstrated live on an open-source kanban board.
AI coding agents are most useful when they can work with your systems — your project tracker, your internal APIs, your data. The Model Context Protocol (MCP) has become the standard way to make that connection, but most examples stop at toy tools. The interesting questions start with existing applications: how do you expose a real API to an agent without handing over broad credentials or unrestricted access?
This talk shows the full path in Python: wrapping an existing REST API into a smaller MCP tool surface with the official mcp SDK and FastAPI, issuing scoped tokens so an agent sees only the projects it should, and connecting agents to the server locally (same machine, containers on a shared network) or remotely. Along the way we cover the design decisions that matter: which operations become tools, read-only by omission, and where the access boundary belongs.
Hi, my name is Harald and I'm a passionate Python developer interested in development, DevOps and AI. I'm currently located in Austria working as a Senior Software Developer and Python Technical Leader for Anexia.
I maintain VibePod, an open-source CLI for running AI coding agents in isolated containers, along with other Open Source Projects. I speak regularly at conferences (PyCon DE & PyData, Grazer Linuxtage, WeAreDevelopers World Congress) and local meetups, teach as a part-time lecturer at universities in Klagenfurt, and write Articles and Tutorials on my blog.