What is the Model Context Protocol (MCP)? A plain-English guide

The Model Context Protocol (MCP) is an open standard for connecting AI applications such as Claude, ChatGPT, Gemini, Cursor and VS Code to outside tools and data. You build an MCP server once for a system, such as a CRM or database, and any MCP-compatible AI app can discover and use its tools, data and prompts.

An AI model on its own only knows what it was trained on and what you paste into the chat. To be useful at work it needs your systems: the CRM, the booking calendar, the shared drive, the database. MCP gives it one standard way to reach them. In December 2025 Anthropic reported more than 10,000 active public MCP servers, and said MCP had been adopted by ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code. That is why it has become the default plumbing for agentic AI.

The official MCP documentation puts it simply: "Think of MCP like a USB-C port for AI applications." It is a good analogy, with one caveat this guide comes back to. Plugging in is easy; deciding what you are willing to plug in, and with what access, is the part that matters for a UK business handling customer data.

  • MCP is an open standard for connecting AI applications to tools and data. It sits in front of your APIs rather than replacing them.
  • An AI app (the host) creates one client per MCP server; each server exposes tools, resources and prompts.
  • Anthropic open-sourced MCP on 25 November 2024; since December 2025 it sits within the Agentic AI Foundation under the Linux Foundation.
  • The current specification, 2026-07-28, made the protocol stateless and deprecated sampling, roots and logging.
  • Every server acts with the access you give it. The ICO expects least-privilege access, and UK GDPR still applies.
Key takeaways

What is the Model Context Protocol?

The Model Context Protocol is an open-source standard that defines how an AI application discovers what an external system can do, reads data from it and calls its actions, using JSON-RPC 2.0 messages. One server per system replaces a separate custom integration for every AI product.

Anthropic announced and open-sourced MCP on 25 November 2024, shipping pre-built servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer, and naming Block and Apollo as early adopters. On 9 December 2025 Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, where it joined goose by Block and AGENTS.md by OpenAI as founding projects. Anthropic said the project's governance model would remain unchanged. For a business, the point is that MCP is not tied to one AI vendor, so a server you build today can serve several AI products.

MCP one year after launch
  • 10,000+ active public MCP servers (December 2025)
  • 97M+ monthly SDK downloads across Python and TypeScript
  • 75+ connectors in Claude's directory, powered by MCP
  • Close to 2,000 entries in the official MCP Registry (November 2025)

Sources: Anthropic, Donating the Model Context Protocol and establishing the Agentic AI Foundation, 2025; Model Context Protocol blog, One Year of MCP, 2025

How does MCP work?

MCP connects three participants: a host (the AI app), a client inside it for each connection, and a server that wraps your system. The host asks each server what it offers, the model chooses a tool, and the server carries out the action and returns the result.

PartWhat it isExample
HostThe AI application the person usesClaude Code, Claude Desktop, VS Code, Cursor, ChatGPT
ClientA connector inside the host that holds one connection to one serverCreated automatically by the host for each server you add
ServerA program that provides tools and context to the AIA server for your CRM, database, file store or a SaaS product such as Sentry
The three participants in MCP

Source: Model Context Protocol, Architecture overview, 2026

  1. The host starts one MCP client per server you add
  2. The client lists the server's tools with tools/list
  3. The host gives the model each tool's name and description
  4. The model picks a tool; the client sends tools/call
  5. The server acts on your system and returns the result
How one MCP tool call flows from host to server

Source: Model Context Protocol, Architecture overview, 2026

Servers run in one of two ways. A local server runs on the same machine as the AI app and talks over the stdio transport, which suits developer tools and files on your computer. A remote server runs as a web service over the Streamable HTTP transport and typically serves many clients, which suits shared business systems that many people, and many AI apps, connect to. The architecture overview notes that MCP recommends OAuth for obtaining authentication tokens on remote servers.

What can an MCP server provide?

An MCP server can expose three kinds of capability, called primitives: tools the AI can call, resources it can read and prompts that structure a task. Most business servers start with a few well-named tools.

  • Tools are actions, such as "search customers", "create invoice" or "run this report". Each has a name, a description and a JSON Schema for its inputs.
  • Resources are data the AI can read for context, such as file contents, database records or a schema.
  • Prompts are reusable templates, such as a standard way to summarise a support ticket.

Servers can also ask the user for input or confirmation part-way through a task, a feature called elicitation. That is useful when an action needs a human decision, such as approving a refund. The original 2024 specification also warned that "Tools represent arbitrary code execution and must be treated with appropriate caution", which is still the right mindset.

How has the MCP specification changed since 2024?

MCP versions are named by date (YYYY-MM-DD), marking the last backwards-incompatible change. The current revision, 2026-07-28, removed the start-up handshake and protocol-level sessions so that every request stands alone, and deprecated sampling, roots and logging.

What each MCP specification revision supports
FeatureNov 2024Mar 2025Jun 2025Nov 2025Jul 2026
Tools, resources, promptsYesYesYesYesYes
OAuth 2.1 authorisationNoYesYesYesYes
Streamable HTTP transportNoYesYesYesYes
Read-only and destructive hintsNoYesYesYesYes
Elicitation (ask the user)NoNoYesYesYes
Structured tool outputNoNoYesYesYes
Long-running tasksNoNoNoPartial, experimentalPartial, extension
Stateless requestsNoNoNoNoYes
SamplingYesYesYesYesDeprecated

Sources: MCP specification 2024-11-05; changelogs for 2025-03-26, 2025-06-18, 2025-11-25 and 2026-07-28, Model Context Protocol

Two practical consequences follow. Older tutorials that describe the initialize handshake or Mcp-Session-Id header describe revisions up to 2025-11-25, so check the specification date before you copy code. And deprecated features normally stay in the specification for at least twelve months under the project's deprecation policy, so existing servers do not break overnight, but new ones should not adopt them.

What is the difference between MCP and an API?

An API is how software talks to a system; an MCP server usually sits in front of one or more APIs and describes them in a form any AI application can discover and call at runtime. MCP does not replace your APIs. It packages them for AI.

QuestionPlain APIFunction callingMCP server
Who uses itDevelopers writing codeOne app talking to one model providerAny MCP-compatible AI app
How capabilities are foundA developer reads the docsTools are written into each appDiscovered at runtime with tools/list
Reuse across AI productsIntegration rebuilt per productTied to the app that defines itBuild once, use from many hosts
Access controlWhatever the API usesHandled by your appOAuth-based framework for remote servers
MCP vs a plain API vs function calling

Source: Model Context Protocol, Architecture overview, 2026

Function calling and MCP work together rather than compete. The host still hands MCP tools to the model as tool definitions; MCP standardises where those definitions come from, so you write them once on the server instead of in every app. If you already have a well-documented API, an MCP server is often a thin, deliberate layer on top: it chooses which endpoints the AI may use and adds descriptions the model can understand.

Which AI tools support MCP?

Support is now broad: Claude, ChatGPT and Codex, Gemini, Microsoft Copilot, Cursor and VS Code all connect to MCP servers. The milestones below come from each organisation's own announcements.

DateMilestone
25 Nov 2024Anthropic open-sources MCP with servers for Google Drive, Slack, GitHub, Git and Postgres
Mar 2025Revision 2025-03-26 adds OAuth 2.1 authorisation and Streamable HTTP
19 May 2025Microsoft announces native MCP support in Windows 11, in private developer preview
Sep 2025The official MCP Registry launches in preview
9 Dec 2025MCP moves to the Agentic AI Foundation under the Linux Foundation
10 Dec 2025Google releases fully managed remote MCP servers, starting with Maps, BigQuery, Compute Engine and GKE
Jul 2026Revision 2026-07-28 makes MCP stateless
MCP adoption milestones

Sources: Anthropic, 2024; Windows Developer Blog, 2025; MCP blog, 2025; Anthropic, 2025; Google Cloud, 2025; MCP changelog, 2026

OpenAI's developer documentation describes MCP as "becoming the industry standard" and covers making a remote server available in ChatGPT and Codex, in ChatGPT deep research and company knowledge, and through the API. The practical value is the same everywhere: build one MCP server for your booking system, and staff can reach it from whichever AI tool they prefer. If you are choosing between the two main coding agents, our comparison of Codex vs Claude Code covers how each handles MCP.

How do you use MCP in Claude Code?

In Claude Code you add a server with the claude mcp add command, either pointing at a remote server's URL with --transport http or giving the command that starts a local server. Each server has a scope that decides who sees it, as the Claude Code MCP documentation sets out.

Claude Code MCP scopes compared
QuestionLocalProjectUser
The default scopeYesNoNo
Shared with your teamNoYes, via .mcp.jsonNo
Available in all your projectsNoNoYes
Checked into version controlNoYesNo
Where it is stored~/.claude.json.mcp.json in the repo~/.claude.json

Source: Anthropic, Claude Code documentation: MCP, checked November 2026

Project scope is the one to use for team tools, because the configuration is reviewed and versioned like the rest of the code. Local scope suits servers with credentials you do not want in version control. Anthropic's documentation is explicit: "Verify you trust each server before connecting it." We cover the rest of a team setup, including skills, hooks and subagents, in Claude Code for business, and our Claude Code agency service sets this up for teams.

Is MCP secure?

MCP is a protocol, so it is as secure as the servers you build and connect. The specification's security best practices set hard rules for the most common mistakes, and prompt injection is the risk no protocol can remove for you.

  • No token passthrough. "MCP servers MUST NOT accept any tokens that were not explicitly issued for the MCP server." A server must not simply forward a user's token to another system.
  • Per-client consent. Proxy servers that sit in front of third-party APIs must get consent for each client, to avoid a "confused deputy" acting for someone without permission.
  • Local servers run code. A client offering one-click local setup must show the exact command before running it, because a local server runs with the same privileges as the AI app.
  • Minimal scopes. Start with low-risk read access and step up only when a privileged tool is first used.

Prompt injection is the remaining risk. An AI that reads emails, web pages or documents through MCP can meet text that tries to give it instructions, and Anthropic's Claude Code documentation warns that servers fetching external content expose you to it. The defence is design: give each tool the narrowest access it needs, keep destructive actions behind a confirmation, and log every call. It is the same discipline as our security checklist for AI-built apps.

How do UK data protection rules apply to MCP?

If an AI tool reads or changes personal data through MCP, that is processing under UK GDPR, and MCP does not change who is responsible. The ICO's 2026 Tech Futures report on agentic AI says controller-processor obligations remain applicable to organisations deploying agents, and that "Organisations should not give agentic AI systems access to information just because it might be useful in the future."

UK and international guidance for connected AI
  • 18 countries' agencies endorsed the UK-developed Guidelines for secure AI system development (2023)
  • 4 life-cycle areas they cover: secure design, development, deployment, and operation and maintenance
  • 10 foundation model developers made or committed to data protection changes after ICO scrutiny (October 2026)

Sources: NCSC, UK develops new global guidelines for AI security, 2023; NCSC, Guidelines for secure AI system development, 2023; ICO, scrutiny extends to AI agents, 2026

In practice that means an MCP server should expose only the records a task needs, act with each user's own permissions, and keep an audit log you can show a regulator. The ICO describes this access management as equivalent to the principle of least privilege. The NCSC guidelines add supply chain security, so treat third-party MCP servers like any other software supplier. This matters most in regulated sectors such as law firms and accountancy practices, where client confidentiality is the whole business.

When should a business build its own MCP server?

Use an existing server when the vendor or a source you trust publishes one for the system you want. Build your own when the system is internal, or when you need to control exactly what the AI can see and do.

  • The system is internal, such as your own database, CRM or line-of-business app.
  • You want AI tools to see only a safe slice of a system, for example read-only reporting views rather than the whole database.
  • You need to enforce business rules, such as approval limits or which records each role may touch, in one place.
  • Several teams or AI products need the same access, and you want to build it once.

A well-designed MCP server is small: a handful of clearly named tools, strict inputs, the user's own permissions and an audit log. That is usually a better first step than giving an agent broad access and hoping its instructions are enough. For what agents do with these connections, see AI agent examples for UK small businesses. If you would rather not build it alone, our MCP server development team designs and builds servers, our AI agent development service builds the agents that use them, and AI integration consulting helps you decide which systems to connect first.

Circlemount is based at St James House in Telford, Shropshire, where Marcus (founder) and a team of 12 specialists work. Our pages for Telford, Shropshire, the West Midlands and Manchester explain how we work with businesses in each area.

Frequently asked questions

What does MCP stand for in AI?

MCP stands for Model Context Protocol. It is an open standard that lets AI applications connect to external tools and data in a consistent way. An MCP server describes what a system can do, and any compatible AI app, such as Claude, ChatGPT, Gemini, Cursor or VS Code, can discover and use it.

What is an MCP server?

An MCP server is a program that wraps a system, such as a CRM, database or file store, and exposes it to AI apps as named tools, readable resources and reusable prompts. It can run locally on your computer over stdio or remotely as a web service over Streamable HTTP, usually protected with OAuth.

Who owns the Model Context Protocol?

Anthropic created and open-sourced MCP in November 2024. In December 2025 it donated the protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. The project is open source, so any company can build MCP servers or clients without a licence fee.

What is the difference between MCP and an API?

An API is how software talks to a system. An MCP server usually sits in front of one or more APIs and describes them as named tools with descriptions and defined inputs, which any compatible AI app can discover at runtime. MCP does not replace your APIs; it packages them for AI.

How do I add an MCP server to Claude Code?

Use the claude mcp add command, either with a remote server's URL and the HTTP transport, or with the command that starts a local server. Choose a scope: local for yourself in one project, project to share it through a committed .mcp.json file, or user for all your projects.

Is it safe to connect MCP servers to business data in the UK?

It can be, with the right design. UK GDPR still applies, so expose only the data a task needs, use OAuth so each user keeps their own permissions, log every tool call and vet third-party servers like any supplier. The ICO and NCSC both stress least privilege and supply chain security.

Does ChatGPT support MCP?

Yes. OpenAI's developer documentation covers making remote MCP servers available in ChatGPT and Codex, in deep research and company knowledge, and through the API. The same server can therefore be used from ChatGPT, Codex, Claude, Cursor and VS Code, which is the main benefit of a shared standard.

Sources

Checked on 16 November 2026. Product features and figures change quickly, so check the current source before you rely on a detail.

Need your systems connected to AI, safely?

Circlemount builds MCP servers that give Claude, ChatGPT and your coding agents the access they need to your CRM, database and internal tools, and no more. Join the Circle for a free business analysis and proposal.