What is MCP?
Author: Domeny.ForSale editorial team Published
In brief: MCP (Model Context Protocol) is an open standard that allows AI applications to use external data and tools in a consistent way. It enables an assistant to read company documentation, check a task in a project management system or invoke a specific operation, for example. MCP is not an AI model and does not itself grant access to all resources — its capabilities depend on the connected services, configuration and permissions.
What is MCP used for, and what problem does it solve?
A language model does not inherently have access to files on a computer or up-to-date data in a company’s systems. That access is provided by the application in which the model runs, through appropriate integrations. Without a common standard, each integration may require a separate way of describing tools, passing parameters and receiving results.
MCP standardises this communication. It defines how an AI application can discover which functions an external system provides and how to use them. The standard was initiated by Anthropic, but it is not limited to a single model or provider.
It can be compared to a common socket for integrations: compatible components are easier to connect without designing the entire connection from scratch. This is a simplification, however. MCP support does not mean that every application supports every feature of the protocol or that a connection will work without configuration.
How do the application, MCP client and MCP server work together?
The MCP architecture has three basic components:
- Host, meaning the AI application used by a person, such as a desktop assistant or code editor. It manages interaction with the model and the rules for using integrations.
- MCP client, meaning a component running within the host that maintains a connection to a specific MCP server.
- MCP server, meaning a program that provides specific data or functions. It can run locally on a computer or as a remote service.
A server can provide tools — functions that perform operations, resources — data to read, and prompt templates — ready-made instruction structures for use in the application, among other things. Not every server offers all these elements.
A typical sequence works like this: the application discovers the available tools, the model proposes using one of them, and the application handles the call according to its own rules. The server performs the operation and returns a result, which may be passed to the model as context for preparing a response. User consent may be part of this process, but you should not assume that every call always requires manual approval.
Example: an assistant checks the team’s tasks
Suppose you want to ask: “Which tasks related to the new website are overdue, and who is responsible for them?” The AI application is connected to an MCP server that supports the company’s project management system. The server provides a task search tool, and its account has read access to the selected project.
The assistant can search for tasks, retrieve deadlines and the people responsible, and then prepare an overview. It uses data returned by the system rather than relying solely on the model’s prior knowledge. However, it may still misinterpret a date or status, so the result should allow you to identify the specific tasks.
The instruction “Move all deadlines back by a week” requires a different capability: writing changes. If the server does not provide such a tool or the account lacks the appropriate permissions, the operation should not be available. For tools that modify data, a sensible safeguard is to display the planned changes and obtain confirmation before proceeding.
How does MCP differ from APIs and RAG?
An API, or application programming interface, allows programs to communicate with a service. MCP does not replace all APIs. An MCP server often calls the API of a project management system, shop or database itself, while exposing selected functions to the AI application in a standard format.
RAG is a way of preparing responses in which the application first searches for relevant material and passes it to the model as context. MCP can provide access to a search engine or document repository used in this process. However, it does not itself define the quality of the search, how documents are split up or the accuracy of the response.
These solutions can therefore work together: an MCP server uses an API to search for documents, and the application uses the results in a RAG process. Connecting MCP does not in itself train the model on company data.
When should you deploy MCP, and how can you limit the risks?
MCP makes sense when an assistant needs regular access to several sources of information or tools, and the application you use supports the standard. For a single, simple connection, a direct API integration may require less work. The choice should be based on the available functions, maintenance costs and access control, not the acronym alone.
Before enabling an integration:
- Check the server’s provenance. A local server is a program that runs on your device; you should not install an unfamiliar package without verifying it.
- Grant minimal permissions. Start with read access to one folder or project rather than access to the entire account.
- Establish the data flow. Check what content is sent to the MCP server and the model provider, and what retention policies apply.
- Secure write operations. Require approval for deleting data, sending messages and making other significant changes.
Content retrieved by tools may contain malicious instructions that attempt to influence the assistant. This threat is called prompt injection; MCP does not eliminate it on its own.
Confirm the supported features and authentication methods in the official documentation for the application and the chosen server. Run your first test using non-confidential data and a read-only account — check both the accuracy of the responses and what the integration actually gains access to.
See also: What is Cloudflare?
Read also
