Sources reviewed: November 17, 2026.
Before you begin
An assistant can write a convincing answer without knowing the actual status of an order. To answer with current information, it needs an authorized way to query the system that holds it. MCP helps standardize that connection.
MCP stands for Model Context Protocol. It is an open standard for connecting AI applications to external sources and tools. It is neither another model nor a replacement for your business software: it defines a common interface through which a compatible application can discover and use capabilities exposed by a server. Official MCP introduction.
How it works, without the jargon
The application you work in acts as the host. Inside it, an MCP client communicates with an MCP server. That server offers specific capabilities, such as accessing documents or querying a system through its API. It can run locally or as a remote service. MCP structures the exchange; it does not determine how the model reasons or how the application uses the context. Official architecture overview.
One request, two different paths
The result returns to the application, which prepares an answer for the person.
Reading: check the status → show the information and its source.
Acting: request a change → check permissions and approval → execute → verify the result.
This is a conceptual diagram. It does not represent an available NOVVOR integration or an architecture suitable for every situation.
What a server can provide
Three concepts help you understand a technical proposal:
- Tools: functions with defined inputs and outputs. Some retrieve information; others can change data or initiate actions.
- Resources: information the application can read and incorporate as context.
- Prompts: reusable instruction templates for working with tools and resources.
Not every server or client supports the same features. “MCP-compatible” does not mean universal access to any system. Check the specific capabilities, supported version and terms of use. MCP server concepts.
A technical detail that matters: the protocol version
This guide references revision 2026-07-28. In that revision, each request carries its version and capabilities in metadata. The initialize handshake and the Mcp-Session-Id header in Streamable HTTP have been removed. The protocol no longer maintains its own sessions. This does not mean a business process cannot have state: a server can use explicit handles passed as tool arguments. Official revision changes.
Older implementations may follow a different flow. Before integrating, verify the versions the client and server actually support. Do not assume these changes apply to every connector already installed. Versioning and compatibility.
A business example, step by step
Imagine a business receiving the question, “What is the status of my order?” For a pilot, we suggest separating four stages:
1. Identify the request. If there is not enough identifying information, ask for it rather than guessing.
2. Query only what is authorized. The tool returns the status, last-updated date and a record reference. It does not need to return the customer's entire profile.
3. Prepare an answer. The application distinguishes verified information from missing information. If the source is unavailable, it says so rather than inventing a delivery date.
4. Review and complete. A person can review the draft. Sending it or changing the order is a separate operation with its own authorization.
For this example, we would measure the accuracy of the reported status, unanswered requests, human intervention and total time. These are suggested pilot evaluation criteria, not results achieved for a client.
MCP, APIs and document retrieval are different things
An API is the interface a system provides to other software. An MCP server can use it to expose an operation to an AI application. Adding MCP does not remove the need to understand the API, its permissions or its limits.
Retrieving relevant documents is also different from executing an action. A project may need document search, system integration or both. Before choosing an architecture, clarify which information changes, what must be queried in real time and which operations must remain out of scope. This is a design distinction, not a requirement to use MCP.
Security: a connection is not authorization
The official security guidance covers risks including token misuse, requests to internal destinations and consent issues between clients and services. Evaluate each server and its provenance, minimize permissions and review how the parties authenticate. Enabling a connector is not enough to assume it inherits all your business safeguards. Official security best practices.
For a project, we recommend documenting:
- Which person or service requests the operation.
- Which data it may read and which data is excluded.
- Which actions require human approval.
- What happens when an error occurs, records conflict or a source is unavailable.
- What evidence verifies the result without retaining unnecessary information.
When MCP is worth evaluating
MCP may be useful when you need to connect an AI application to several reusable capabilities. For a deterministic task, a conventional integration may be simpler. Our recommendation is to compare alternatives using the same test case, rather than adopt a protocol because it is popular.
A good initial scope has a frequent task, an identifiable source, a responsible owner and a verifiable outcome. Starting with read-only queries lets you learn about quality and access before considering changes to systems.
Questions for the first conversation
Which task do we want to improve? Where does the correct information live? Who can access it? What must the AI not do? How will we know it has failed? Who will review exceptions?
Answering these questions is more useful than starting with a list of models. MCP may be one part of the solution; the process and its controls remain the starting point.
A real example: Microsoft Learn
Microsoft maintains an MCP server for accessing its official documentation. Compatible clients can search content, retrieve complete articles and find code samples. This is a concrete example of making knowledge available within a workflow. It does not demonstrate that an agent has permission to administer a business or modify its systems. Microsoft Learn MCP Server.
Frequently asked questions
Does MCP replace an API?
No. It can give an AI application a common way to discover and use operations that already rely on an API. The API's limits, permissions and data quality still matter when designing the project.
Does connecting through MCP let AI change any data?
No. Connection and authorization are separate decisions. For a pilot, we recommend starting with narrowly scoped queries and separately evaluating any operation that writes, sends or deletes information.
Do I need MCP to bring AI into my business?
Not necessarily. Before adding components, compare a direct integration with an MCP-based solution on the same task. Choose the option that meets the objective with clear controls and a reasonable maintenance burden.
Start with a concrete task
A productive conversation starts with a task worth improving, not “let's connect everything.” Define the expected result, the correct source and the limits on action. That foundation helps you evaluate where AI adds value and where a conventional process remains the better choice.
Continue reading: AI agents or automation: what does your business actually need?


