The short answer
MCP, the Model Context Protocol, is an open standard that lets AI applications connect to outside data and tools in one consistent way. An MCP server wraps a service, such as a CRM or a drive, and offers it to any compatible AI app as tools, resources and prompts. Anthropic introduced MCP in 2024 and has since donated it to the Agentic AI Foundation, a Linux Foundation project. The current specification is dated 28 July 2026, and Claude, ChatGPT plugins, Zapier and Make all support it.
| Full name | Model Context Protocol |
|---|---|
| Who looks after it | Agentic AI Foundation, under the Linux Foundation |
| Current specification | Version 2026-07-28 |
| Three parts | Host (the AI app), client (connector inside it), server (the service) |
| What a server offers | Tools, resources and prompts |
| Where you meet it | Claude custom connectors, ChatGPT plugins, Zapier MCP, Make MCP server |
What MCP is, in plain English
Before MCP, every AI app needed its own custom integration for every service it wanted to talk to. A connector built for one assistant did not work in another. MCP replaces that with one shared plug. TechCrunch's AI glossary calls it "a USB-C port for AI": build an MCP server for your service once, and any AI app that speaks MCP can use it.
The official specification at modelcontextprotocol.io describes MCP as an open protocol for connecting AI applications to external data sources and tools. It says it took inspiration from the Language Server Protocol, the standard that lets code editors support many programming languages. Anthropic introduced and open-sourced MCP on 25 November 2024. Its post of 22 October 2026 confirms that it has since donated MCP to the Agentic AI Foundation, a project under the Linux Foundation that Anthropic co-founded with Block and OpenAI, so the standard is no longer run by a single company.
How MCP works: hosts, clients and servers
The specification names three roles, and once you know them the rest makes sense:
- Host: the AI application you use, such as Claude, ChatGPT or Cursor.
- Client: the connector inside the host that holds the connection to one server.
- Server: the service that offers data and actions, for example a server for your project tracker or your company wiki.
A server can offer three kinds of things. Tools are actions the model can run, such as "create ticket" or "search orders". Resources are data and context, such as a file or a record. Prompts are ready-made templates and workflows for users. In the other direction, the host can let a server ask the user for missing information, which the specification calls elicitation. Messages travel in a standard format called JSON-RPC 2.0. The July 2026 specification also lists optional extensions, including Tasks for long-running jobs, Skills over MCP for structured instructions and MCP Apps for charts and forms shown inside a conversation.
Where you will meet MCP at work
- Claude. Custom connectors use remote MCP servers. On Pro and Max you add one from Customize, then Connectors, then + Add and Add custom connector, and paste the server URL. Free accounts can add one custom connector; on Team and Enterprise an owner adds it in Organization settings first. See our Claude guide.
- ChatGPT. Plugins bundle skills, apps and MCP servers. When custom GPTs retire on 11 December 2026, OpenAI says GPT custom actions do not transfer and should be replaced by a supported app or a custom MCP server. See what is a custom GPT.
- Automation platforms. Zapier MCP lets Claude, ChatGPT, Cursor and other clients take actions in more than 9,000 apps, with each tool call using two Zapier tasks. Make's MCP server lets those clients run your on-demand scenarios. n8n asks for explicit approval before an MCP connection is saved. See Zapier, Make and n8n.
- Other assistants. Mistral AI says the 100+ connectors in its assistant Vibe are MCP compatible, and Google said on 8 October 2026 that its Gemini agent for Gemini Enterprise can connect to any MCP server inside or outside your company network.
Why it matters at work
MCP is what turns a chatbot into something closer to a colleague with access to your systems. Instead of copying data out of your CRM and pasting it into a chat, the assistant can look it up and act on it. For teams, that means one integration serves every AI tool you use, so switching assistants no longer means rebuilding your connections. Our guide to automating tasks with AI shows where this fits.
It also means new risk. The specification itself says tools represent arbitrary code execution and must be treated with caution, that hosts must get explicit user consent before invoking any tool, and that descriptions of tool behaviour should be treated as untrusted unless the server is trusted. Anthropic's help page adds that custom connectors can reach services it has not verified, that malicious servers can hide prompt injections, and that you should choose Allow always only for tools you trust. Our guide on using AI at work safely covers the policy side.
An example you can try
The quickest safe way to see MCP working is with a read-only server in Claude:
- Pick a service you already use that publishes an official remote MCP server URL in its own documentation. Do not use URLs from forums or unknown sites.
- In Claude, open Customize, then Connectors, choose + Add, then Add custom connector. Give it a name, paste the URL, check the authentication settings and select Add.
- When Claude asks to use a tool, read the request before approving it. Do not choose Allow always until you trust the server and the tool to run unsupervised, which is Anthropic's own advice.
- Ask a question that needs the service, using the first prompt below, and compare the answer with what you see in the service itself.
Common misconceptions about MCP
- "MCP is an AI model." It is not. It is a protocol, a set of rules for how AI apps and services talk to each other.
- "MCP is an Anthropic product." Anthropic created it, but it is now governed by the Agentic AI Foundation under the Linux Foundation, and OpenAI, Google, Mistral AI and many automation tools support it.
- "MCP replaces APIs." Usually an MCP server sits on top of a service's existing API and presents it in a form AI apps understand.
- "If it connects through MCP, it is safe." The protocol cannot enforce security on its own, the specification says so plainly. Safety depends on the server you trust and the permissions you grant.
Related terms
- AI agent: the kind of assistant that uses MCP tools to act.
- AI API: the programmable access MCP servers often wrap.
- Prompt injection: the main security risk with connected tools.
- RAG: another way to give a model your data.
- Computer use: AI that clicks through apps when no connector exists.
Ready to build something? See how to use AI agents, compare Zapier vs Make or Make vs n8n, and pick from the best AI for automation.
Prompts to copy
Check a connector before you rely on it
Using the [connector name] connector only, list the last [10] [records, for example open deals] with their [owner, stage and value]. Then tell me exactly which tool calls you made and what each returned, so I can compare it with the source system. Do not create, edit or delete anything.
You get: A read-only answer plus an audit trail of tool calls
Review an MCP server's tools for risk
Here is the list of tools and their descriptions from an MCP server we are considering: [paste tool list]. Sort them into read-only and write actions. For each write action, explain what could go wrong if the AI called it by mistake, and recommend which tools we should disable or keep on approval only.
You get: A permission plan for a new connector