MCP explained simply: what the Model Context Protocol is and what it means for you

MCP in plain words

On its own, an AI assistant knows nothing about your calendar, your notes or your company's files. For it to access them, it has to be connected to those services. In the past, each connection had to be built separately.

MCP is meant to make this simpler. The project website describes MCP as an open standard for connecting AI applications such as Claude or ChatGPT to data sources, tools and workflows. It compares it to a USB-C port: just as one standard plug connects different devices, MCP connects AI programs to other systems.

As examples, the project website mentions an assistant that accesses your Google Calendar and Notion notes, or a company chatbot that queries several databases.

How it works technically, in brief

MCP has two sides. On one side is the AI application, such as Claude or ChatGPT. On the other is a so-called MCP server. This is a small program provided by a service that tells the AI what it can do there: for example read appointments, run a search or create an entry.

Such servers can run on the internet, in which case they are called remote servers. In the apps you usually connect to them through so-called connectors or connected apps. But they can also run directly on your own computer if you download and install them.

So the terms do not mean the same thing: MCP is the standard itself. An MCP server is a program that implements this standard. A connector is the connection in your AI app that can, for example, use an MCP server.

Anthropic introduced MCP in November 2024. In December 2025, Anthropic handed the standard over to the Agentic AI Foundation, a fund under the umbrella of the non-profit Linux Foundation. It was co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, Amazon Web Services, Cloudflare and Bloomberg. At the time, Anthropic spoke of more than 10,000 active public MCP servers.

Who supports MCP today (as of October 2026)

Claude: according to Anthropic, custom connectors via remote MCP servers are available in Claude, Cowork and Claude Desktop on the Free, Pro, Max, Team and Enterprise plans. On the free plan you get one custom connector. On Team and Enterprise plans, only certain administrators may add connectors.

ChatGPT: according to OpenAI, full MCP support, including actions that change things, is rolling out in beta for the Business, Enterprise and Edu business plans. On Pro, MCP servers with read-only permissions can be connected in developer mode. The German pricing page also lists developer mode as a beta for Plus; exactly what it covers there is not clear from the information available.

Google: according to Google's help pages, custom MCP connections in the Gemini app are only available in the US, for personal Google accounts and for users aged 18 and over. For developers, Google has offered MCP servers for its own services such as Google Maps and BigQuery since December 2025.

Programming tools such as Visual Studio Code and Cursor also support MCP, according to the project website.

What this means for you as an ordinary user

Most of the time you will not notice MCP at all. You see a list of connectors or connected apps in your AI assistant, click "Connect" and sign in to the service. Whether MCP is behind it makes little difference to how you use it.

According to the project website, MCP is meant to let AI assistants access your data and act on your behalf when needed. That is handy when the AI is supposed to find free slots in your calendar or summarise notes, for example.

Adding your own MCP servers by address or installing them on your own computer, on the other hand, is more for technically experienced users. Our assessment: for most people in Germany today, it is enough to use the ready-made connectors from the big providers and to pay attention to permissions.

What to watch out for on security

All three big providers explicitly warn about risks. Anthropic advises connecting only servers from organisations you trust, checking the requested permissions carefully and refusing unnecessary access. According to Anthropic, malicious servers can contain hidden instructions intended to make Claude take unwanted actions. Experts call this prompt injection. Providers can also change how their tools behave without notice. Anthropic says you should only choose the option to always allow an action for services you fully trust.

OpenAI writes that unsafe or untrusted MCP servers increase the risk of such attacks. Google points out that it neither controls nor secures third-party MCP servers and that your data there is subject to that provider's privacy rules. According to Google, Gemini asks before write actions but can make mistakes. Some services also request more data than they need.

The MCP project's own security guidance also warns that servers running on your own computer can have direct access to your system. So only download such programs from sources you trust.