What is an MCP server, and when does a mid-sized firm need one?

Short answer: An MCP (Model Context Protocol) server is a small piece of software that lets AI (Artificial Intelligence) assistants such as Claude, ChatGPT and Gemini use your product's data and actions through one open standard. You need one when your customers or staff already ask an AI assistant to do work that your software could do for them.

Where MCP came from

Anthropic released the Model Context Protocol as an open standard on 25 November 2024. OpenAI adopted it in March 2025 and Google DeepMind followed in April 2025. In December 2025 Anthropic gave the protocol to the Agentic AI Foundation, a fund under the Linux Foundation co-founded with Block and OpenAI, so no single AI company controls it.

The current version of the specification is dated 28 July 2026. It added a stateless core, which makes servers easier to run at scale, and tightened the rules on authorisation.

For a buyer, the point is simple. Before MCP, connecting your product to each AI assistant meant a separate integration for each one. With MCP you build one server and every assistant that speaks the protocol can use it.

How an MCP server works, in plain terms

There are three parts.

1. The host. The AI application your user is working in, such as Claude or ChatGPT.
2. The client. A connector inside the host that talks to one server.
3. The server. The part you build. It sits in front of your existing API (Application Programming Interface) and describes what the AI is allowed to do.

The server offers three kinds of thing:

• Tools: actions the AI can take, such as "create an invoice" or "look up an order".
• Resources: data the AI can read, such as a product catalogue or a policy document.
• Prompts: ready-made instructions for common jobs, such as "summarise this customer's account".

When a user asks the assistant to do something, the assistant reads the tool descriptions, picks the right one, asks the user to sign in if needed, and calls your server. Your server checks the user's permissions and then calls your API as normal.

When a mid-sized firm needs one

You probably need an MCP server if two or more of these are true:

• Your customers already copy data out of your product and paste it into an AI assistant.
• You sell to businesses, and their staff use Claude, ChatGPT, Microsoft Copilot or Gemini at work.
• A competitor has listed a connector in an AI assistant's directory and you have not.
• You have a working, documented API.
• Your own team wants AI agents to work across your internal systems, such as a CRM (Customer Relationship Management system), a finance system and a ticketing tool.

You probably do not need one yet if your product has no API, if all you want is a chatbot on your website, or if none of your users work inside AI assistants.

What it costs and how long it takes in the UK

The cost depends on how many actions the AI needs, how sign-in works and how sensitive the data is. A read-only server with a handful of tools is at the low end. A server that lets the AI write data, handles per-user permissions and passes a security review sits higher.

At RudraX, connector and MCP server projects are priced from £6,500 on a fixed price, with typical delivery in 3 to 8 weeks. Building in-house usually takes longer because the team has to learn the protocol, tool design and AI testing at the same time as their day jobs. You can get a rough figure for your own project from our AI development cost calculator.

Security: the part most guides skip

An MCP server gives an AI model a way into your systems, so it needs the same care as any public API, plus a few extra checks.

• Sign-in with OAuth 2.1. OAuth (Open Authorization, the standard sign-in protocol) lets each user sign in with their own account, so the AI only sees what that person is allowed to see. Never share one master key across all users.
• Read first, write later. Launch with read-only tools. Add actions that change data only once you have seen how the AI uses the read tools, and ask the user to confirm each write.
• Guard against prompt injection. Text inside a document or web page can try to trick the model into calling a tool it should not. Keep tool permissions narrow and log every call.
• Do not pass tokens through. The MCP security guidance forbids a server from forwarding a user's token to another service unchecked, because it creates a "confused deputy" risk.
• Use your identity provider. Since June 2026 the protocol supports enterprise-managed authorisation, so firms on Microsoft Entra or Okta can control which staff can use which servers.

This matters more in the UK this month. The ICO (Information Commissioner's Office) is running a call for evidence on agentic AI that closes on 20 November 2026, and its questions on security, transparency and accountability apply directly to MCP servers that handle personal data.

How to start

1. List five to ten jobs your customers would hand to an AI assistant if they could.
2. Check your API covers those jobs, with proper authentication.
3. Design the tools with clear names and descriptions. This step decides whether the AI picks the right action.
4. Build and test in more than one assistant. Claude and ChatGPT read tool descriptions differently.
5. Review security before launch, then list the server in the assistants' directories.

If you want a second opinion on whether your product is ready, read how we approach AI connector and MCP server development, see our work for SaaS (Software as a Service) clients, or book a free 30-minute call.