MCP · 7 min read
What is MCP? A plain-English guide for UK businesses
By the Hot Rocket team
MCP, the Model Context Protocol, is an open standard that lets AI assistants use other software through a common interface. A business builds one MCP server for a system, such as a CRM or a document store, and any MCP-capable assistant can then use the tools it exposes, within the permissions the business sets.
The short version
MCP stands for Model Context Protocol. It is an open standard for connecting AI assistants to other software. Instead of writing a one-off integration for every pairing of an assistant and a system, you build one MCP server for your system. Any assistant that speaks MCP, called an MCP client, can then discover what the server offers and use it.
Anthropic published MCP as an open standard on 25 November 2024.1 In December 2025 it was placed under the Agentic AI Foundation, a directed fund at the Linux Foundation, so no single vendor controls it.2 The current revision of the specification is dated 28 July 2026.3
What an MCP server actually exposes
An MCP server publishes three kinds of thing:
| Primitive | What it is | A business example |
|---|---|---|
| Tools | Actions the assistant can call, with typed inputs | find_customer, create_quote_draft, check_stock |
| Resources | Read-only data the assistant can pull into context | A price list, a policy document, a product catalogue |
| Prompts | Reusable, named instructions a user can pick | "Summarise this account for a renewal call" |
The assistant sees each tool's name, description and input schema. When a person asks for something, the model decides which tool to call, the client sends the request, and the server runs it, using the server's own permissions, not the model's imagination.
Plain-English test. If you can describe a job as "look something up in system X, then do Y", an MCP server is probably the cleanest way to let an assistant do it safely.
Why it matters to a business, not just to developers
One integration, many assistants. Your staff might use Claude, ChatGPT, Microsoft Copilot or an in-house agent. A single MCP server for your CRM can serve all of them, so you are not locked into one vendor's plug-in format.
Control sits with you. The server decides what is exposed, to whom, and with what limits. You can make a tool read-only, require approval before anything is written, restrict it to certain records, and log every call. The model never gets your database password.
It is auditable. Every tool call is a discrete request with inputs and outputs. That is far easier to review than an assistant that has been handed a broad API key.
What changed in the 2026 specification
The 28 July 2026 revision made the protocol stateless. The old initialize handshake and session header were removed. Each request now carries its own protocol version and client details, and servers must answer a server/discover request that advertises what they support.4 For a business, the practical effect is that MCP servers behave like ordinary web APIs. They sit behind a load balancer, scale horizontally and restart without losing anything, which makes them simpler and cheaper to run.
What MCP is not
- It is not an AI model. It is plumbing between models and your systems.
- It is not a security layer by itself. The server has to implement authentication, authorisation and logging properly. A badly built server can expose too much.
- It does not make an assistant correct. The model can still call the wrong tool or misread a result. That is why anything consequential should keep a person in the loop.
When to build one, and when not to
Build an MCP server when:
- People already ask an assistant questions whose answers live in one of your systems.
- The same system needs to be reachable from more than one assistant or agent.
- You need the access to be permissioned and logged, for example because personal data is involved.
Do not bother when a simple scheduled automation would do. A nightly export or a webhook into your accounts package does not need an AI model in the middle.
What a first MCP project looks like
A sensible first project is read-only. Expose three to five lookup tools against one system, for example "find a customer", "list open jobs" and "get an invoice's status". Connect the server to the assistant your team already uses and log every call. Once people trust the answers, add write actions one at a time, each with an approval step.
Try ours
This website runs a small, read-only MCP server at /mcp. It lets an assistant list our services, read one in detail and answer common questions. It is a working example of the pattern described here.
Sources
-
Anthropic, "Introducing the Model Context Protocol", 25 November 2024. anthropic.com/news/model-context-protocol ↩
-
Linux Foundation, "Linux Foundation announces the formation of the Agentic AI Foundation", December 2025. linuxfoundation.org ↩
-
Model Context Protocol specification, revision 2026-07-28. modelcontextprotocol.io/specification/2026-07-28 ↩
-
"Key changes" for revision 2026-07-28, which lists changes since revision 2025-11-25. github.com/modelcontextprotocol ↩
T+99Next step
Talk it through with us
A free 30-minute first chat, with no obligation. Tell us what is getting in the way; we tell you honestly what we would do, roughly what it would take, and whether it is worth doing at all.