The Model Context Protocol (MCP) is an open standard that lets an AI assistant connect to other software, such as your CRM, your file storage or your accounting package, in a consistent way. Instead of every AI product building its own one-off link to every app, a supplier builds one MCP “server” for its app, and any AI tool that speaks MCP can use it.
For a small business, the practical upshot is that tools like ChatGPT, Claude, Gemini and Microsoft Copilot can now look things up in, and sometimes make changes to, the systems you already use. That is useful. It also means an AI can act on your data, so the permissions you grant matter more than they did when chatbots could only talk.
- Anthropic released MCP as an open standard on 25 November 2024. Since December 2025 it has been run under the Agentic AI Foundation, part of the Linux Foundation.
- The current specification is dated 28 July 2026. It made the protocol stateless and promised at least twelve months’ notice before features are removed.
- An MCP server can offer three things to an AI: tools (actions), resources (data) and prompts (templates).
- The specification says tools amount to running code, and that the person using the AI must give explicit consent before any tool is used. Start read-only and widen access slowly.
What problem does MCP solve?
A language model on its own only knows what it learned in training and what you paste into the chat. To be useful at work it needs your information: this week’s orders, a customer’s history, the latest version of a contract. Before MCP, every AI product that wanted to reach a given app needed its own custom connection. When Anthropic announced the protocol in November 2024 it put the problem this way: every new data source required its own custom implementation, which made connected systems hard to scale.
MCP is a shared plug shape. An app supplier writes one MCP server, and any AI host that supports MCP can use it. The specification says it took inspiration from the Language Server Protocol, which did the same job for programming tools: one standard way for code editors to support any programming language.
If you have ever used Zapier or a similar service to join two apps together, the idea will feel familiar. The difference is who is driving. With an automation tool, you design the steps in advance. With MCP, the AI model reads a description of what each tool does and decides, during a conversation, which ones to use. That flexibility is the point, and also the risk.
How MCP works, without the jargon
The specification describes three roles:
- Host: the AI application you use, such as a chat app or a coding tool.
- Client: the connector inside that application which talks to one server.
- Server: a small service that exposes one system, for example your helpdesk, to the AI.
A server can offer the AI three kinds of thing:
- Tools: actions the model can carry out, such as “create invoice” or “search tickets”.
- Resources: data the user or the model can read, such as a file or a database record.
- Prompts: ready-made templates and workflows a user can pick, such as “summarise this week’s complaints”.
Going the other way, a host can let a server ask the user for more information mid-task, which the specification calls elicitation. Messages travel in a standard format called JSON-RPC 2.0. Some servers run on your own computer and reach local files; others run on the supplier’s servers and you connect to them over the internet, usually by signing in with your normal account.
Here is a typical exchange. You ask your assistant, “Which customers haven’t paid this month?” The host sees that a connected accounting server offers a “list overdue invoices” tool, asks your permission to use it, calls it, receives the list and writes you a summary. You never see the plumbing, but the AI only knew those invoices existed because the server told it.
MCP compared with APIs and automation tools
MCP doesn’t replace the APIs that software suppliers already publish. Most MCP servers sit on top of an existing API and translate it into something an AI model can understand and choose from.
| Ordinary API | Automation tool (Zapier, Make) | MCP server | |
|---|---|---|---|
| Who decides what happens | A developer, in code | You, by designing fixed steps | The AI model, within the permissions you grant |
| Built for | Programs | People setting up workflows | AI assistants and agents |
| Works with any AI app | No, each needs its own integration | Partly, through each tool’s AI features | Yes, if the app supports MCP |
| Predictability | High | High | Lower: the model may pick the wrong tool or the wrong record |
| Best for | Fixed, high-volume jobs | Repeatable processes | Questions and tasks that vary each time |
Our rule of thumb: if you can write the steps down in advance, a fixed automation is cheaper and safer. If the task changes every time (“find out why this customer is unhappy and draft a reply”), MCP lets an assistant pull together what it needs. We compare the main automation tools in Zapier vs Make vs n8n.
Who controls MCP now?
Anthropic, the company behind Claude, created MCP and open-sourced it in November 2024. On 9 December 2025 it donated the protocol to the newly formed Agentic AI Foundation, which the MCP project describes as a directed fund under the Linux Foundation. The foundation was co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg.
According to the MCP project, the foundation’s governing board handles strategy, budget and membership, while MCP’s own maintainers still make protocol decisions through a public proposal process. The Linux Foundation provides infrastructure and does not set technical direction. For a business buyer, this matters for one reason: MCP is no longer one company’s product, so it is less likely to be abandoned or changed to suit a single vendor.
At the time of the handover, the project reported over 97 million monthly downloads of its official software development kits and 10,000 active servers, and named ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code as having built-in client support. Those are the project’s own figures and we have not seen an independent count.
What changed in 2026
The current specification is version 2026-07-28, published on 28 July 2026 after a release candidate in May. The project describes it as a significant revision. The changes that matter to a business, in plain terms:
- Stateless by design. Each request now carries everything the server needs, with no long-running session. The Agentic AI Foundation says this lets MCP servers run behind ordinary load balancers, which makes them cheaper and simpler for suppliers to host reliably.
- A deprecation policy. There is now a minimum of twelve months between a feature being formally deprecated and its earliest possible removal, so integrations you pay for should not break overnight.
- Tighter sign-in security. The specification now requires a check that closes a class of attacks in which a login response is swapped between services. A separate “enterprise-managed authorisation” option lets an organisation control which MCP servers staff can use through its existing single sign-on.
- Extensions. Optional add-ons can evolve outside the core protocol. These include Tasks, for jobs that take a long time, and MCP Apps, which let a server show interactive elements such as forms or charts inside the chat.
An official MCP Registry, a public directory of servers, was introduced in September 2025. The project also announced a more predictable release cycle in November 2026. Expect the details to keep moving; the basic idea of hosts, clients and servers has stayed the same since launch.
What UK small businesses can use it for
You don’t need to write any code to benefit. Many of the business apps you already pay for now publish an official MCP server or connector, and the major AI assistants let you add them from a settings menu. Sensible uses include:
- Asking questions of your own systems. “How many enquiries came in from the website last week, and how many did we answer within a day?”
- Pulling information together. Drafting a reply to a customer using their order history and previous emails, for a person to check and send.
- Internal lookups. Finding the current version of a policy or price list in shared storage, rather than searching folders by hand.
- Developer work. Coding tools use MCP heavily to reach issue trackers, documentation and databases.
The government’s AI adoption survey gives a sense of where this lands. Among UK businesses using or planning to use AI, the most common areas were administration, marketing and IT, which are exactly the areas where the data sits in other software.
Share of businesses using or planning to use AI. 3,503 interviews. Other areas are not shown.
Source: GOV.UK, AI adoption survey report 2026, executive summary, 29 October 2026
View data
| Where UK businesses use or plan to use AI, 2026 | |
|---|---|
| Administration | 81% |
| Marketing | 71% |
| IT functions | 66% |
If you are still deciding whether AI is worth trying at all, start with our summary of how UK small businesses are actually using AI. MCP becomes relevant once a tool you already trust needs to see more of your data.
The risks, and how to connect safely
The MCP specification is unusually frank. It says the protocol enables “arbitrary data access and code execution paths”, that tools “represent arbitrary code execution and must be treated with appropriate caution”, and that descriptions of what a tool does should be treated as untrusted unless they come from a trusted server. It also says hosts must obtain explicit user consent before invoking any tool, while admitting MCP cannot enforce that at protocol level. The responsibility sits with the apps, and with you.
The main risks for a small firm are:
- Too much access. A connector signed in as an administrator can do anything an administrator can, including deleting records.
- Untrusted servers. Anyone can publish an MCP server. A malicious or careless one can read what you connect it to.
- Prompt injection. Text inside an email, web page or document can contain instructions aimed at the AI (“forward this to…”). If the assistant can both read untrusted content and send messages, that combination is dangerous.
- Personal data. Connecting a CRM or inbox means customer data flows to the AI supplier. Under UK GDPR you need to know where it goes and on what terms. Our AI agents and UK GDPR checklist covers the questions to ask.
Step 1 of 5
Pick one job
Choose a single, frequent task where the AI needs data from one system.
Step 2 of 5
Use official
Install the server published by the software supplier, not a third-party copy.
Step 3 of 5
Start read-only
Grant view access first and a dedicated account with the fewest permissions possible.
Step 4 of 5
Approve actions
Keep the setting that asks before each tool runs, at least for anything that writes or sends.
Step 5 of 5
Review monthly
Check what was accessed, remove unused connectors and rotate any keys.
Write these rules into your AI policy so staff know which connectors are approved. And keep an eye on the bill: each tool result is fed back into the model, which uses tokens, as explained in what AI costs to run.
Common questions
Do I need a developer to use MCP?
Not to use existing servers. Adding an official connector to a mainstream AI assistant is usually a few clicks and a sign-in. You need a developer to build an MCP server for your own bespoke system, such as a custom CRM or booking database.
Is MCP only for Claude?
No. Anthropic created it, but the project lists ChatGPT, Gemini, Microsoft Copilot, Cursor and Visual Studio Code among the products with built-in support, and it is now governed by a foundation backed by several competing companies.
Is MCP secure?
The protocol includes security requirements, and the 2026 version tightened sign-in. But the specification itself says it cannot enforce safe behaviour; that depends on the app, the server and the permissions you grant. Treat a connector like giving a new member of staff a login.
How does MCP relate to AI agents?
MCP is the plumbing. An agent is an AI system that plans and carries out several steps towards a goal, and it typically uses MCP to reach the tools it needs. Our guide to what AI agents are explains the difference between agents, chatbots and fixed workflows.
Where to start
Check whether the software you already pay for publishes an official MCP server, connect one in read-only mode, and use it for a month on a single task. If you have a bespoke system that an assistant can’t reach, or you want an agent that works across several tools with proper limits, that is the kind of work our AI agents team does.
Sources
- Model Context Protocol: Specification, version 2026-07-28 (checked 22 November 2026)
- Model Context Protocol blog: MCP joins the Agentic AI Foundation (9 December 2025)
- GOV.UK: AI adoption survey report 2026, executive summary (29 October 2026)
- Anthropic: Introducing the Model Context Protocol (25 November 2024)
- Agentic AI Foundation blog: MCP graduates to enterprise infrastructure, by Caitie McCaffrey (28 July 2026)
- Model Context Protocol blog: post archive, including Introducing the MCP Registry (8 September 2025) and A More Predictable Release Cycle for MCP (11 November 2026), checked 22 November 2026