16 November 2026, 7 min read
What is an MCP server, and should your company knowledge have one?
An MCP server lets AI tools read and act on your systems through one standard. How it works, and what to check before you expose knowledge.
An MCP server is a small piece of software that lets an AI tool use something outside itself: a database, a file store, a calendar or a company knowledge base. The "MCP" stands for Model Context Protocol, and the server is the side that holds the data or the ability. If you have heard the phrase a lot lately and are not sure what it adds to a normal API, this post explains it plainly, then gets to the question a company should care about: should your knowledge have one?
Where MCP came from
Anthropic open-sourced the Model Context Protocol on 25 November 2024. Its stated problem was that AI models were cut off from the places data lives, and every new data source needed its own custom integration. A single open standard was meant to replace that pile of one-offs.
The official documentation describes MCP as an open-source standard for connecting AI applications to external systems, and compares it to a USB-C port: one connector, many devices. Assistants such as Claude and ChatGPT, and development tools such as Visual Studio Code and Cursor, are listed as supporting it.
How an MCP server works
There are three roles.
- Host. The AI application, for example Claude Desktop or an editor.
- Client. A component inside the host that keeps one connection to one server.
- Server. The program that provides context or abilities.
The architecture overview explains that the host creates one client per server. Messages use JSON-RPC 2.0. A local server typically talks over standard input and output, while a remote one uses Streamable HTTP, where bearer tokens, API keys and OAuth are all options.
A server can offer three kinds of thing:
| Primitive | What it is | Example for a knowledge base |
|---|---|---|
| Tools | Functions the AI can call | Search records, read a record, propose a change |
| Resources | Data the AI can read | A policy page, a team's record |
| Prompts | Reusable templates | "Summarise this team's projects" |
The client asks the server what it has (tools/list), the model chooses, and the client calls the tool (tools/call). Because the listing is dynamic, a server can show different tools to different people. That is useful and, as we will see, important.
How it differs from a REST API
A REST API is a contract for developers. Someone reads the docs, writes code and decides when to call what. An MCP server is a contract an AI model can read for itself. It advertises its tools, their descriptions and their input schemas, and the model decides in context which to use.
In practice the two sit side by side. The same system often has a REST API for scripts and an MCP server for assistants, backed by the same data and the same rules. Neither replaces the other. If a product only has one, it is not a problem, but you should know which one your AI tool is using.
What the protocol does not do
The protocol defines how a client and server talk. It does not decide what a server returns to a given person. The architecture overview is explicit that the scope is the exchange of context, and that what an application does with the model or the context is outside it.
That leaves the interesting question to the server author. Who is calling, and what are they allowed to see? A server that answers every request with the same credential gives every user the same view. If that credential is broad, you have built a very convenient way to read everything.
Standardising the connection made it easy to plug AI into company data. It did not make it safe. Safety is a property of the server.
Should your company knowledge have an MCP server?
If you want people to ask Claude, ChatGPT or Cursor questions about your company and get real answers, then yes, in some form. Pasting documents into a chat does not scale, goes stale and bypasses every permission you set.
But "have one" is the beginning of the question. These are the properties that decide whether it is a good one.
It authenticates a person. Calls arrive with an identity, not just a shared key. Per-person tokens or OAuth let the server know who is asking.
It applies that person's access on every call. The tools list, the search results, the individual reads and any export all pass through the same permission check. If the server can return something the person could not open in the product, that is a bug.
Unauthorised records look absent. A search should not say "three results hidden". A read of a record you cannot see should look the same as a read of one that does not exist.
Writes are scoped and reviewable. If tools can change data, they should be limited to what the person may change, and changes that contradict a human's entry should wait for review.
It is observable. You should be able to see which tool was called, by whom, for whom and when.
It is small. A handful of well-described tools is easier to secure and easier for a model to use well than fifty overlapping ones.
Local and remote servers
The distinction between local and remote servers changes the security picture, so it is worth a short section.
A local server runs on the same machine as the AI application. A typical example is a server that reads files from a folder you choose. The trust boundary is your own computer, and it usually serves one user.
A remote server runs somewhere else and serves many people over the network. A company knowledge base is almost always this kind. It needs real authentication, because the connection crosses a network and the same server is answering for many different people. The architecture overview notes that remote servers typically serve many clients, and that OAuth is the recommended way to obtain tokens.
For company knowledge, prefer a remote server that signs each person in, rather than a local script that holds a long-lived key on every laptop. A key in a config file is easy to copy, hard to revoke selectively and tied to no one in particular.
What to put behind it
An MCP server is a doorway. What matters is the room behind it. If the data is a pile of loosely organised pages, the server can only search the pile. If the data is typed records with owners and relations, it can answer structured questions: who owns this, who reports to whom, which policies apply to which team.
That is the reason we pair the access argument with a structure argument. We wrote more on the first in how to give AI agents your company's knowledge without leaking what they shouldn't see, and on the second in corporate wikis vs company knowledge bases.
MCP in both directions
There is a second use that gets less attention. A knowledge base can be an MCP server for AI tools, and it can also read from MCP servers. If your CRM, ticketing system or internal service already has one, the knowledge base can use it as a source instead of needing a bespoke integration. That keeps the amount of glue code small: one standard on the way in, one on the way out.
A short test before you connect anything
- Connect the server to an AI tool as a user with limited access.
- Ask for something that user may not see. Check the answer treats it as missing.
- Ask the same question as a user with full access. Check the answer differs.
- Remove access from the first user and repeat. Check it takes effect immediately.
- Look in the logs for each call.
Five minutes of this tells you more than any feature list.
How Contextely approaches this
Contextely exposes your company knowledge over MCP, and also over a CLI and a REST API, all backed by the same rules. Each connection sees exactly what the person it acts for can see, so Claude, ChatGPT, Cursor and custom agents get different answers for different people.
It can also read from other systems: you can connect any MCP server as a source, alongside file upload and a read-only Postgres connection. Permissions are enforced for every path, and if an agent contradicts something a person wrote, it files a proposal instead of overwriting it.
The short version
An MCP server is the standard way for AI tools to reach your systems. Your company knowledge should have one, as long as it knows who is asking, applies their access on every call and keeps humans in charge of what is true. The protocol will connect things. You still have to decide what it is allowed to connect.
Questions
What is an MCP server?
A program that gives AI applications context and abilities through the Model Context Protocol, an open standard. It can expose tools the AI can call, resources it can read and prompt templates. Claude, ChatGPT and Cursor can all act as clients to such servers.
Is an MCP server the same as an API?
They overlap, but an MCP server is described in a form an AI model can discover and use directly, with named tools and typed inputs. A REST API is written for developers. Many products offer both over the same underlying data.
Should my company knowledge have an MCP server?
If you want AI tools to answer from it, yes, but the server must enforce permissions. The protocol standardises the connection, not who is allowed to see what. That part is up to whoever builds the server.
Is it safe to connect an MCP server to an AI tool?
It depends on the server. Prefer ones that authenticate each user, apply that user's permissions to every call, and log what happened. Be wary of servers that share one broad credential across everyone.
Do I need to write my own MCP server?
Usually not. If your knowledge base already offers one, connect to it. Write your own only for internal systems that have none, and design its permissions before its tools.