Model Context Protocol (MCP) – the standard that connects AI to the outside world

Definition

MCP (Model Context Protocol) is an open standard protocol that lets AI systems communicate with external tools, data sources and services in a standardised way – rather than each integration having to be built from scratch.

What is MCP?

Model Context Protocol (MCP) is an open standard protocol launched by Anthropic in November 2024. It defines how AI applications – such as chat assistants, coding assistants and agents – can connect to and communicate with external systems: databases, APIs, files, tools and services.

The easiest way to understand MCP is to think of it as USB for AI integrations. Before USB, every peripheral needed its own connector and its own driver. USB solved that with a universal standard. MCP does the same for AI: instead of building a custom integration for every combination of AI system and external tool, MCP defines a shared protocol language that everyone can use.

The problem MCP solves

Without a standard, you get what is usually called an M×N problem. If you have M different AI applications (chat assistants, RAG systems, agents) and N different external systems (GitHub, Slack, databases, Google Drive), you could need up to M×N unique integrations. That leads to duplicated work, inconsistent implementation and systems that are hard to maintain and scale.

MCP turns M×N into M+N. System vendors build an MCP server once. AI application developers implement an MCP client once. After that, every client can connect to every server without any further integration work.

How MCP works – the architecture

MCP is built on a client–server architecture with three roles:

• Host – The AI application the user interacts with, for example Claude Desktop, an IDE assistant or a custom agent. The host contains one or more MCP clients.
• Client – The component inside the host that maintains the connection to a specific MCP server. Each client has a 1:1 connection to a server.
• Server – The external program that exposes data or functionality via MCP. An MCP server for GitHub lets the AI read code and create pull requests. A server for Postgres lets the AI query a database in natural language.

Communication runs over JSON-RPC 2.0, the same messaging standard used by the Language Server Protocol (LSP) – the technology that lets code editors offer intelligent code suggestions whatever the programming language.

What an MCP server can expose

MCP defines three types of primitive that a server can offer:

• Tools (model-controlled) – Functions that the AI model itself decides to call: fetch a file, run a search, send a message. Think of it as function calling with a shared standard.
• Resources (application-controlled) – Data sources that the application provides as context: a document, a database row, a repo. Resources are information retrieval without side effects.
• Prompts (user-controlled) – Predefined templates and workflows that the user can invoke directly, for example "summarise the latest meeting minutes from Notion".

Why MCP became an industry standard so quickly

Since its launch in November 2024, MCP has spread remarkably quickly. In March 2025, OpenAI announced that it was adopting MCP and integrating it into the ChatGPT desktop app. Google DeepMind has done the same. Thousands of MCP servers are now available – for GitHub, Slack, Asana, Postgres, Figma, Webflow, Ahrefs and many more.

In December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF) – part of the Linux Foundation and co-founded by Anthropic, Block and OpenAI. That cements MCP as an open, neutral standard that no single company owns.

Claude.ai itself uses MCP – it can connect to Google Calendar, Gmail, Ahrefs and Webflow via MCP servers, for example.

MCP and AI agents

MCP is a fundamental building block of modern AI agents. An agent needs to be able to interact with the outside world – read data, carry out actions, communicate with other systems – and MCP standardises exactly those integrations. Without a protocol like MCP, every agent implementation has to build its own connections from scratch, which fragments the ecosystem and makes agents hard to reuse and combine.

For companies that want to build or deploy AI agents, understanding MCP is now essential. An agent that connects via MCP to your CRM, your document library and your project management platform can carry out tasks that previously required human coordination – with your documentation and your processes as context.

Security aspects

With MCP's power comes responsibility for security. The protocol enables arbitrary code execution and data access, so clients must obtain explicit user consent before calling tools. It is crucial that MCP servers are given only the permissions they actually need, and that a human stays in control of high-impact actions that could have negative consequences.

MCP in practice – a few examples

• Coding assistants: IDEs such as Cursor and Zed use MCP to give AI coding assistants direct access to the whole project context – files, git history, documentation.
• Business automation: An HR agent can use MCP to read employee data from the HR system, create drafts in Google Docs and send reminders in Slack – without a human intermediary coordinating the steps.
• Data analysis: An agent connected to a Postgres server via MCP can answer business questions asked in natural language, straight from your database.

If you would like to understand how MCP works and how the right AI agents could improve your organisation, contact us.