MCP
What is MCP and how does it work?
MCP (Model Context Protocol) is an open standard that allows artificial intelligence (AI) applications to connect to external data sources, tools, and systems. It defines how an AI application can discover available functions, utilize them, and receive results.
It standardizes the transmission of information and queries, but does not itself determine how the AI application should use the data received.
MCP is not a separate AI model or chatbot. It is a communication protocol between an AI application and the systems connected to it.
Context
Why did MCP emerge?
AI applications need a way to access information that is not in the model's training data: company documents, current system records, or user account information. This requires integrations. When different AI applications connect to each source in different ways, the number of separately developed and maintained solutions grows.
MCP offers a common standard for these integrations. System functions can be provided through an MCP server, which is accessed by AI applications that support this protocol. This creates the opportunity to reuse the same integration, instead of creating a completely separate connection method for each AI environment.
It does not eliminate integration work, but standardizes part of it.
Flow
How does MCP work?
MCP works on a client-server principle. When using a function provided through MCP, the process can be divided into four steps.
The AI application receives information about available tools
The MCP server provides their names, purpose, and the data that needs to be specified in the query.
A tool suitable for the task is selected
The AI model can suggest using it, and the application forwards the query according to the rules set within it.
The MCP server processes the query
It performs the intended action itself or contacts another system.
The result is returned to the AI application
It can use the received information to prepare a response or continue the task.
MCP itself defines the rules of this exchange - what tools to present to the model and how to use their results is decided by the AI application developers. For data-changing or other significant actions, human approval may be required: invoking an MCP tool does not automatically mean that the action should be performed without the user's knowledge.
Server
What is an MCP server?
An MCP server is a program that provides access to specific data or functions through the MCP protocol. It can enable reading documents, performing searches, querying databases, or using functions from other systems. The server developer defines what capabilities will be available.
The word "server" here refers to a program, not necessarily a separate computer. An MCP server can run on a user's computer, in corporate infrastructure, or as a remote service. Its location and connection method depend on the specific solution.
Client
What is an MCP client?
An MCP client is a component of an AI application that communicates with a specific MCP server. It sends requests, receives responses, and passes them to the application used by the person.
It's important to distinguish between the AI application itself and its MCP client. In MCP architecture, the user's application is called a host. It manages the user experience and can have multiple MCP clients for connecting to different servers. This allows a single AI environment to use capabilities from multiple connected systems.
Capabilities
What capabilities does an MCP server provide to an AI application?
MCP defines three main ways in which a server can provide its capabilities: tools, data resources, and prompt templates.
Tools
Functions that can be called: perform a search, retrieve system data, execute a calculation, or modify a record.
Resources
Information that an AI application can read and use as context: document content, data structures, or other provided material.
Prompts
Reusable query or instruction templates that help perform specific tasks.
Not every MCP server has to provide all three capabilities. One may be designed only for tools, another for reading information or working with prepared templates.
Comparison
How does MCP differ from an API?
An Application Programming Interface (API) defines how one application can use the functions of another application. Through an API, you can retrieve data, transmit it, or initiate predefined actions in the system.
MCP is designed to standardize the presentation of data and functions to AI applications. It encompasses not only function invocation but also a way to discover what tools are available, what they are intended for, and what data is required to use them.
MCP and APIs can work together. An MCP server can use an existing system API and perform the action requested by the AI application through it. Therefore, MCP doesn't necessarily replace existing API integrations - it can provide an additional, AI-optimized access method.
Comparison
How does MCP differ from RAG?
RAG (Retrieval-Augmented Generation) - a method for generating AI responses by leveraging additionally retrieved information. Before generating a response, material related to the query is selected from documents or other sources and presented to the model. This does not require retraining the model.
MCP addresses a different question: how an AI application communicates with a data source or tool. It is not itself an information retrieval method and does not determine how to select the most appropriate documents for a response.
These solutions can be combined. For example, document search required for a RAG solution can be provided as an MCP tool. However, RAG can operate without MCP, and MCP can be used for tasks that are not related to document-based response generation at all.
Security
Is MCP secure?
MCP support alone does not guarantee a secure solution. The protocol defines communication principles, but access control, user consent, and data protection must be implemented by the developers of specific applications and integrations.
First and foremost, it is important to assess what permissions the integration receives: access to one document group and the ability to modify all system data pose different risks. We evaluate:
- What permissions the integration receives
- Whether permissions are limited according to the task being performed
- Whether their compliance is verified on the system side
- Whether the selected MCP servers are trustworthy
- Whether data transmitted through them is verified
- Whether additional confirmation is required for sensitive actions
Simply instructing an AI model not to use certain data is not enough. Another risk is malicious instructions in external data or tool descriptions, called prompt injection: these may attempt to induce the AI to perform undesirable actions or transmit information.
Frequently asked questions about MCP
MCP integration with business systems
We help evaluate MCP applicability to your systems and plan the necessary integrations.
Discuss MCP integration