Bilatzailea
MCP: Standard protocol for connecting artificial intelligence systems to the world
To do more advanced things than just chatbots, AI systems need to connect to the world to search for information or influence it in other ways. And, depending on the function of the system, it will have to be connected to multiple external systems. The implementation and orchestration of connections and calls for each of these systems greatly complicates their development. To do this, the Model Context Protocol (MCP) has been created, which allows AI models to use any external system in a single standard way.
In the previous issue we explained what RAG systems are. To improve the accuracy of chatbots based on LLM or large language models (especially in highly specialized or current topics), before asking the question to LLM, a search is made (on the Internet, in documents of our company...) Calling an API and passing the first result pages or documents to the LLM saying “Answer this user question by relying on the contents of these documents.” It is noted that in a RAG system there is a layer with the ability to call an API above the LLM.
If the RAG system only needs to perform the search on a collection, it will need to implement an API call. But if you need to consult more than one collection, then calls to more than one API will have to be implemented, and the work will be added and complicated. In addition, you may have to decide which font to consult depending on the question, or depending on the result you will have to see if you need to do more searches, in the same or another font... It can be really complex to program this layer on a RAG system; if the logic is not very superficial, it is often not easy to do this as a computer program.
MCP: standard for chatbots to have access to the world
To facilitate this, the MCP protocol was created. The acronym stands for Model Context Protocol, indicating that it is a protocol for providing context to models. They provide a single universal interface for LLM-based AI systems to query APIs. It was created by the company Anthropic (Claude LLM and the creators of the chatbot) in November 2024, but in December 2025 it entrusted the ownership and future development of the protocol to the Agentic AI Foundation, which is the responsibility of the Linux Foundation. Over time, all chatbot providers (OpenAI, Google, Microsoft...) have moved to use and promote it, and it can be said that it is the standard that everyone uses today.
Thus, the APIs to be used by AA systems currently overlay a layer of an MCP server. The MCP protocol allows you to ask this server what functions it makes accessible, with what parameters and types, and what kind of results it returns. And then you can be told to execute one of these functions with the desired parameters, and the result will be received.
simplifying the orchestration of complex RAG systems
The operation of an RAG system such as the one mentioned above, using the MCP protocol, would be as follows: each available MCP server that is desired to be used is asked for its functions, parameters and results, and then the LLM is called “Answer this user question by searching through these sources, using different functions in each of them.” The LLM then decides on which sources the documents can be useful for the query, tells the system which MCPs to call and with which parameters, and when it receives the result it decides whether it will perform more searches or already has enough information to answer the user.
It is noted that the only advantage of the MCP protocol is not that it saves the work of implementing different APIs. When there are several sources for different types of searches, the task to be performed may be complex. The logic of these can be difficult and/or expensive or impossible to implement in a computer program.
Instead, source selections and other decisions will be left to the LLM, and the RAG system itself becomes something very simple and quite general: It must know how to make calls to the MCP protocol and will initially ask the MCP servers what their functions, parameters, etc. are; then, for each user question, it will only pass to the LLM a generic instruction always the same as the one we have seen, then make the calls requested by the LLM to the MCP servers until it tells the user that it already has the answer, and finally pass the answer to the user. This complicated logic, which would be difficult or impossible to program, will not have to be programmed, and it will be the LLM that is there to do this type of thing that will orchestrate the MCP servers.
Beyond the search
In the examples used in this article to explain what MCP is, all MCP servers have been for searching for information. In fact, complex RAG systems are one of the cases where MCP is useful. We had already seen what RAG systems are and the MCP services used in them are only for search.
But the fact is that the servers that can be made available through the MCP protocol are not only for searches and queries. On the contrary, you can and do MCP servers that can execute all kinds of actions: consult and write email, book flights and hotels, make purchases from an online store, move our robotic warehouse cars, interact with code stores... Today, all types of MCPs are available. And we can make our own to access our documentation, to interact with our computer...
Systems that use MCP servers that offer things beyond the search are AA agents. You’ve probably heard that term lately, haven’t you? It's a very hot topic, no doubt. We will talk about what they are, how they work and their advantages and problems in the next issue.
Buletina
Bidali zure helbide elektronikoa eta jaso asteroko buletina zure sarrera-ontzian
