An MCP server is a service that gives an AI assistant access to specific data and functions, such as a document archive or a database, through the Model Context Protocol (MCP). A company needs one when it wants staff to work with company knowledge inside the AI assistant they already use, without building a bespoke integration for every tool.
What exactly is the Model Context Protocol?
The Model Context Protocol is an open standard that Anthropic introduced in November 2024. Its specification is public and describes how AI applications communicate with external data sources and tools. Technically, MCP is built on JSON-RPC 2.0, a well-established format for messages between programs.
The specification defines three roles:
- Host: the AI application a person works in, for example a chat assistant on the desktop or in the browser.
- Client: the connector inside that application that maintains the link to a server.
- Server: the service that provides data and functions.
An everyday comparison: MCP does for AI assistants what a standard plug socket does for electrical appliances. Any appliance with the right plug works in any socket, and the socket does not need rewiring for each new device. The specification itself draws a similar parallel with the Language Server Protocol, which connects programming languages to development tools in a uniform way.
An MCP server can run locally on one machine or be reachable over a network. The specification defines two standard transports for this: stdio for local processes and Streamable HTTP for servers on a network.
What does an MCP server provide?
An MCP server can offer the assistant up to three kinds of building block. The server’s operator decides which ones to expose.
| Building block | What it is | Example in a company |
|---|---|---|
| Tools | Functions the AI model can call | “Search the project archive for fire safety conditions”, “Open page 14 of file X” |
| Resources | Data and context for the user or model to read | A document, a project list, a record |
| Prompts | Ready-made workflows and message templates for users | “Summarise the status of a project with sources” |
The key point is that the assistant does not get blanket access to all data. It can only use the functions the server offers, and the server decides what it returns.
Why would a company need an MCP server?
An MCP server solves a practical problem: company knowledge sits in archives, line-of-business applications and databases, while AI assistants sit in a chat window. Without a connection, the assistant answers questions about the business from general knowledge, or not at all.
Take a fictional engineering firm, Muster & Partner, with an archive of twenty years of expert reports and project files. With an MCP server for that archive, a project manager can ask her AI assistant what conditions the authority imposed on comparable projects. The assistant calls the server’s search tool, receives the relevant passages with file and page, and writes an answer backed by those sources. The method behind this is retrieval-augmented generation (RAG), covered in our article Searching company knowledge with AI.
That gives a company three advantages:
- No new application. Staff work in an assistant they already know, which keeps training effort low.
- One connection, several assistants. Because MCP is an open standard, the same server can in principle be used with different assistants that support MCP.
- Control stays with the server. Access rights, logging and the choice of what data is released are handled on the server, not in the chat window.
How does MCP compare with other ways of connecting AI to data?
MCP is not the only way to connect AI with company data. The table sets the usual options side by side.
| Option | Advantages | Drawbacks |
|---|---|---|
| Uploading documents to the chat by hand | No technical effort | Only for one-off cases; no permission checks; whole documents go to the provider |
| Own chat application with a direct API integration | Complete design freedom | A separate interface to build, maintain and train people on |
| Vendor-specific plugin | Deep integration with one product | Tied to one vendor; must be rebuilt if you switch |
| MCP server | Open standard; familiar interface; control on the server side | Feature set depends on how much of MCP the assistant supports |
For companies that want to bring their knowledge into existing AI assistants, MCP is currently the option with the lowest lock-in risk in most cases. A bespoke application still makes sense where a process needs a dedicated interface.
What about security and data protection?
The MCP specification sets out explicit principles on security and trust: users should knowingly consent to data access and actions, tools are to be treated as potentially dangerous code execution, and tool descriptions from untrusted sources should be handled with caution. The specification also states that the protocol cannot enforce these principles by itself. That responsibility lies with whoever builds the servers and applications.
For a company, this means:
- Know the data flow. With a cloud-based AI assistant, whatever the server returns goes to the assistant’s provider. A good server therefore releases only the passages an answer needs. Personal data requires a legal basis and a data processing agreement under Article 28 GDPR.
- Sign-in and permissions. The server must know who is asking and return only what that person is allowed to see.
- Trusted servers only. Staff should not connect arbitrary MCP servers from the internet. Approval belongs with IT.
- Limit write access. Read-only tools carry less risk than tools that change data or send emails.
Our security and privacy page explains how we handle these questions.
How do you recognise a usable MCP server for your company?
Before introducing an MCP server, run through these questions:
- Which tools does the server offer, and are they read-only or do they write?
- How do users sign in, and are existing access rights respected?
- What data leaves the server with each request, and who receives it?
- Is there a log of who retrieved which data and when?
- Does the server return source references so that answers can be checked?
- Which AI assistants has the server been tested with?
- Who runs and updates the server, and where does it run?
How does UNID AI Second Brain use MCP?
UNID AI Second Brain is delivered as an MCP connector, so the document archive appears directly in the AI assistant staff already use. The archive, search index and records stay on a controlled server. Only the passages an answer needs are sent to the assistant, each with its file and page. See the Second Brain product page for details and the glossary for definitions. To talk about your archive, use our contact page.