Various LLMs also support the feature, to call up functions (Function Calling). These language models recognize when a user prompt requires the use of an auxiliary tool.
Model Context Protocol relies on this function calling capability, but offers a more standardized way to connect to services and data sources. This makes it possible to connect AI systems with external sources without having to reinvent the wheel or program it every time.
The MCP architecture
Let’s now take a look at the various components of the MCP architecture and how they work together.
- The MCP-Host is the AI-based application that provides the connection to the outside world. When Anthropic introduced MCP, it integrated the protocol into its Claude desktop application, making it one of the first MCP hosts. However, hosts are not limited to LLM-based chatbots. An AI-supported IDE could also serve as a host, for example. A host program includes the core LLM as well as a variety of utilities.
- The MCP-Client is the most important of these aid programs. Every LLM needs a client that is customized to the way it calls tools and processes data. However, all MCP clients provide a standard set of services to their hosts. They search for available servers and report back these services and the parameters required to call them. All of this information is incorporated into the prompt context of the LLM. When the LLM detects a user input that should invoke an available service, it sends a corresponding request via the client to the respective MCP server.
- Everyone MCP-Server is designed to communicate with a data source or external service in a language that they “understand”. It also communicates with the MCP client, acting as a kind of intermediary between the two components.
- MCP client and MCP server communicate with each other in a JSON-based format. The MCP-Transport-Layer converts MCP protocol messages to JSON-RPC format for transmission and converts JSON-RPC messages back at the receiving end. It should be noted that a server can run locally on the same computer as the client or online. In the first case, client-server communication takes place via stdio, in the second case via streamable HTTP.
The following diagram demonstrates how the MCP components work together.
Foundry
The MCP server ecosystem
Although the architecture and its components seem complex at first glance, the modularity, portability and standardization of MCP ensure that developers have an easier life. Because:
- An MCP server can communicate with any AI application that includes a properly implemented MCP client. This means that if you want to make your service available to AI agents, you can create a single server and be assured that it will work with various types of LLMs.
- Conversely, an MCP client must be tailored to a specific host, but can be connected to any properly implemented MCP server. So you don’t have to figure out how to connect a specific LLM to Google Docs, a MySQL database, or a weather forecast service. All you have to do is create an MCP client. This can establish a connection to all kinds of services via MCP servers.
For deeper insights into the MCP Server ecosystem, we recommend visiting the MCP Servers Repository on GitHub. (fm)
