Building plugins and datasets

A plugin is a git repo that teaches the MCP server about a set of datasets. We scope each plugin to a focused data domain (for example Uruguay's energy balance) rather than to a whole open data portal, which tends to grow too general. The Uruguay energy and Brasil catalogs are plugins; yours can be too.

A plugin can describe its tools in two ways, and mix both freely:

• Python tools: plain Python functions. This is the main path, and what we reach for in practice: it handles anything from a simple lookup to databases, APIs and computations, and stays clear as a dataset grows.
• YAML datasets: declare a query in a small .yaml file, no programming required. Only for really simple datasets: see when to use YAML vs. Python.

We prefer Python functions as tools to start. Whatever the style, every tool must follow the same result contract: a text answer for the AI plus structured data (sources, tables, charts) for the UI.

The path to your own plugin

1. Start from the existing catalogs as templates: Uruguay's energy plugin (Python).
2. Write your first tool as a Python function.
3. Test it locally with MCP Inspector.
4. Give your plugin a description and sample questions so the chat shows a nice landing card.
5. Connect it to a server.

Two lessons from the pilots

Two practices proved especially valuable when building a plugin: precompute derived values instead of asking the AI, and give the AI a domain glossary by injecting official definitions into your tools' context.