An MCP server is a small, agreed way for an AI assistant to use specific tools and sources, instead of people pasting information into a chat window. It is a technical standard with a practical point: the assistant can only see and do what you have chosen to expose, and a person can still review what comes back.
MCP stands for Model Context Protocol, an open standard for connecting an AI application to external tools, files and systems. Think of it as a plug: the assistant on one side, and on the other a focused service that can fetch an approved document, look up a record, or run a bounded action you have defined.
This article is a plain-English guide for charities and SMEs. It is not a specification, and it is not a claim that every organisation needs one. It is here to help you ask better questions when a supplier, a developer or a colleague says “we’ll add an MCP server”.
The problem it is trying to solve
Most teams first meet AI as a chat box. That is fine for a quick draft from information you are willing to type in. It works much less well when the useful knowledge already lives in folders, a CRM, a case system or a database, and when staff should not be pasting those records into a public tab.
Without a defined connection, people copy and paste. They also invent slightly different prompts each time, and nobody can see which sources were used. An MCP server is one way to replace that habit with a declared interface: these are the tools, these are the sources, and this is all the assistant is allowed to call.
That is a close cousin of the idea behind Ask Your Data — answers from your own records, with a trail of what was asked — and of the boundaries in private AI for charity data. MCP is not the only way to build that kind of system. It is a standard shape for the connection.
What an MCP server actually is
In practice, an MCP server is a programme that offers a short list of capabilities to an AI application that supports the protocol. Those capabilities might be “search this approved folder”, “fetch this template”, “list open cases the user is allowed to see”, or “create a draft in this format”. The assistant can request a capability. The server performs that one thing, or refuses, and returns a result the assistant can use in its reply.
The important design choice is what you put on that list. A server that can see an entire shared drive, or take actions across a live system with no review, is not automatically a better server. A narrow server, tied to one job, is usually easier to explain, test and supervise.
The protocol does not, by itself, make the arrangement private, UK-hosted, accurate or safe. Those are separate decisions about where the server runs, which accounts can call it, what data it can reach, and who checks the output. For the difference between local and hosted models, see local versus cloud AI models.
Why charities and SMEs might care
A well-designed connection can reduce blank-page work without turning staff into prompt engineers. A fundraiser, case worker or office manager can ask a question or request a first draft, and the assistant can use the approved tool rather than asking them to upload a file they should not be copying around.
It can also make governance easier to describe. Instead of “people sometimes paste things into a chatbot”, you can say “the assistant may call these tools, against these sources, and a person signs off anything that goes out”. That is the kind of sentence trustees and owners can understand. It also gives you somewhere to put logging: which tool was called, by whom, and when.
None of that removes the need for human review. If a draft will be sent, filed, or relied on, a person still checks it. An MCP server is infrastructure for access and consistency. It is not a substitute for judgement, and it will not fix a process you cannot already describe.
What to ask before you agree to one
Which job is this for? If the answer is “so the AI can use all our systems”, ask for one job first.
What tools and sources will it expose? Get a written list, including what it cannot do. “Search the policy folder” is a tool. “Read any mailbox” is a much larger decision.
Where does it run, and where does data go? On equipment you control, or through a hosted service? What is logged, and for how long?
Who is allowed to use it? Named people or roles, not a shared login that nobody owns.
How is the output checked? Especially where a mistake would affect a person, a payment, or a report.
How will we test it? Realistic questions, awkward cases, and an attempt to make it retrieve something it should refuse. The same habits as in can your AI be trusted.
If you cannot get plain answers to those questions, you are not ready to put real records behind the connection. Official guidance such as the NCSC guidelines for secure AI system development and the ICO guidance on AI and data protection is still the reference for those design choices, whether or not MCP is involved.
How to think about a first use
A sensible first MCP server, if you have a reason to use the standard at all, is small: one role, one class of questions, one approved source, and a reviewer. After that has been used in real work, you can decide whether a second tool is justified.
If your actual need is “ask a question of our own records and see where the answer came from”, start by describing that job. The protocol, if it is used, should follow the job, not the other way around. The same is true of any AI process improvement work: map the job, then build around it.
Questions
Do we need an MCP server to use AI at work?
No. Many useful first projects are a bounded assistant, better instructions around a tool you already have, or a small workflow with a human review. MCP is one way to connect an assistant to tools and sources when that connection needs to be explicit.
Is an MCP server the same as a private AI system?
No. It is a connection standard. Privacy, hosting, access control and review are design choices you still have to make.
Is it safe to expose a live case system or CRM through one?
Only with a narrow, agreed set of tools, strict access rules, logging, and a person who checks anything that matters. A wide-open connection to a live system is usually the wrong first step.
If you want help turning one repeating job into a bounded assistant or a clear connection to your own records, book a free 40-minute call by phone or Teams. We’ll send a written summary of next steps.
Book a free call