Short definition
MCP — the Model Context Protocol, an open standard that lets AI applications such as ChatGPT, Claude or Gemini connect to external tools and data sources in a consistent way, through an MCP server that decides which functions are available.
Also known as: Model Context Protocol, MCP server.
MCP stands for Model Context Protocol. It is an open standard that lets an AI assistant such as ChatGPT, Claude or Gemini use the tools and data of other software in a consistent way. Instead of building a separate integration for each assistant, you build one MCP server, and every application that supports MCP can work with it.
How does MCP work?
The specification defines three roles. The host is the application you work in, for example Claude Desktop or ChatGPT. Inside it sits a client, which maintains the connection to one server. The server is the piece of software that offers a system’s functions: your CMS, your calendar, a product catalog or an internal database.
An MCP server can offer three kinds of building blocks:
- tools: actions the model can request, such as
search_productorcreate_draft, each with a name, a description and fixed input fields; - resources: data the application can read, such as a document or a file list;
- prompts: ready-made instructions a user can pick.
The connection runs locally over stdio when the server runs on your own computer, or remotely over Streamable HTTP. For remote servers, the specification describes authorization based on OAuth 2.1. That way a user signs in with their own account, and the assistant only gets the permissions that come with that account.
The model does not run any code itself. It requests a tool, the client sends that request to the server, and the server decides whether and how the action happens. The rest of the work lives in that server: which functions exist, which input is allowed and what comes back.
Who supports MCP?
Anthropic released MCP on November 25, 2024 as an open standard. On December 9, 2025, Anthropic donated the protocol to the Agentic AI Foundation, a new foundation within the Linux Foundation, with OpenAI among the co-founders. The latest version of the specification is dated July 28, 2026.
As of November 2026, the major assistants support MCP, each with its own conditions:
- ChatGPT: according to OpenAI’s help page, custom MCP servers are available for ChatGPT Business and Enterprise/Edu on the web. Only admins and owners publish such a connection for their workspace.
- Claude: custom connectors using remote MCP work on every plan, including Free, although Free is limited to one custom connector. Claude Desktop can also start local MCP servers.
- Gemini: according to Google, you can add a custom app with the URL of an MCP server in the Gemini web app. The Gemini CLI for developers also works with MCP servers.
- Copilot: GitHub Copilot Chat can use MCP servers for extra context. In Microsoft Copilot Studio, an agent supports MCP tools and resources, according to the documentation.
These conditions change quickly. For a concrete project, always check the current documentation of the assistant your team or customers use.
Examples of MCP servers
An MCP server is as broad or as narrow as the task. A server can grant read-only access to Search Console, or offer one function that prepares a draft in a CMS without publishing it. Which CMS platforms already have their own server is covered in MCP servers for CMS platforms.
We build MCP servers ourselves too. One example is Ruokago, our own ordering platform for restaurants. Its staging and demo environment runs a public MCP server with ten tools. On November 21, 2026, we requested the real list with tools/list. Five tools are marked read-only, such as search_menu, check_delivery and get_availability. The others work with the shopping cart. The last step, checkout_link, only returns a link to the restaurant’s checkout. The tool itself never places or pays for an order. The full tool list is in our article on agentic commerce for your online store.
For our own work, we use smaller servers. One of them lets an AI agent check Dutch and English text for spelling. Seppe wrote two articles on dev.to about that server and about a server that searches documentation and keeps track of the source: giving an AI agent a spell checker with MCP and a small docs MCP server.
Is MCP safe?
MCP itself is an agreement about how software talks. Whether it is safe depends on what the server allows. The specification for tools says a person should always be able to step in and deny a tool call. It also warns that a client should not blindly trust a tool’s labels, such as “read-only,” when the server is not trusted.
Three risks deserve attention:
- Overly broad permissions. A server that can do everything makes every model mistake expensive. Start with read-only access and grant write access per task.
- Instructions hidden in data. A product description, an email or a web page can contain text that looks like a command. That is called prompt injection. The Ruokago server therefore tells the model explicitly that menu names and descriptions are data and never instructions.
- Passed-through access. The security best practices call it an anti-pattern when a server simply passes a client’s tokens on to other systems. A server checks for itself who is allowed to do what.
So never let the model calculate or decide amounts, stock levels or permissions. Those come from your own system, and the server checks every input again.
The biggest misconception: AI search does not find your server on its own
A public MCP server does not make your website visible in ChatGPT, Gemini or Perplexity. None of the major assistants looks for MCP servers by itself during an ordinary question. A user or admin adds the server deliberately, or you publish an app through that platform’s own route. That was the outcome of our research in November 2026, and the practical side is covered in the article on agentic commerce.
Visibility in AI answers is still search work: clear content, good structured data and a site crawlers can read. The glossary entries on AI Overviews and generative engine optimization cover that side.
Having an MCP server built?
A custom MCP server makes sense when people who already work with an AI assistant need to use your system directly. Two situations come up often:
- your team: employees work in Claude or ChatGPT and want to look up data from your own software or have a draft prepared, without copying and pasting everything;
- your customers: you offer a service or catalog and want an assistant the customer already uses to prepare the cart, as with Ruokago, while checkout stays on your own site.
Building a server only to be found more easily makes little sense, and neither does building one for a single fixed connection between two systems. A regular API integration or workflow automation is simpler for that.
If you want an assistant that also handles tasks itself, such as updating your website or bundling questions, we can build an AI agent for your business. Other integrations and automations fall under AI and automation. Either way, we decide together with you which permissions the agent gets. Design, code and the review before anything goes live stay human work.