MCP explained without hype or fluff

Summary

Model Context Protocol, like most protocols, solves the M ⨯ N integration problem by turning it into an M + N integration problem. An AI client application that speaks this protocol does not have to figure out how to fetch data or take actions specific to a platform. MCP may or may not make your AI smarter, or improve your product, but it will reduce the friction to integrate against other applications that already support MCP. This may or may not be important to you. The protocol specifies MCP Servers, that generally connect to data sources and expose tools specific to it. Then there are MCP clients, which are a part of AI applications. They can connect to any MCP Server, typically through a configuration that specifies how to connect to or run the server. The servers, more commonly implemented than clients, may expose: Tools that the LLM can call, eg, fetch_file for a filesystem or send_mail for a mail client integration. Prompts, which are reusable templates of instructions or multi-step conversations for the LLM, that are intended to be user-controlled. Resources that are exposed via URIs; it’s up to the client application’s design to decide how these are fetched or used. Sampling, which allows servers to request LLM completions on the client application, which is useful for agentic patterns and running context-aware inference without needing to receive all the contextual data from the client. There are a few more functions and nuances to servers, but these are what broadly stood out to me. Most servers that I have seen or used mostly just expose tool calls. A tiny concrete example: an MCP server for Open Data access I wrote a tiny MCP server to expose actions to take on CKAN, an open source data management system that’s used by Governments and other organisations to publish open datasets. CKAN has a web interface that links to these tagged datasets, which are usually semi-structured (CSVs, XLS) or totally unstructured (PDF reports and papers). This is not parti...