What is MCP, and how does it plug AI into the tools you already use?

MCP, short for Model Context Protocol, is an open standard that lets an AI assistant such as Claude or ChatGPT connect to the apps where your work lives, so it can read your calendar, search your Drive or update a Notion page. You switch it on through a settings screen, and you choose which actions the assistant may take on its own and which ones need your click first. I based this page on the official MCP documentation and specification, Anthropic's launch post, and the help pages of each AI app named below, all of which I opened on 15 October 2026, plus the connectors I run in my own work.

What is MCP in plain words?

MCP is a shared plug format between AI apps and other software. The official MCP site compares it to "a USB-C port for AI applications": one standard shape, so any assistant that supports it can plug into any tool that has an MCP server.

Before MCP, every link between an AI app and a tool had to be built by hand. When Anthropic released MCP on 25 November 2024, it described the problem this way: "Every new data source requires its own custom implementation." With a shared standard, Notion builds one MCP server and it works in every assistant that speaks MCP.

Two words come up whenever you read about MCP. The tool side is the server: Gmail, Notion or your ad account, wrapped so an assistant can use it. The assistant side is the client. A server gives the assistant tools, which are actions such as creating a calendar event, and resources, which are data the assistant can read, such as a file.

MCP is no longer one company's format. The governance page shows the project set up under LF Projects, LLC, with code and specification under the Apache 2.0 licence. The current specification is dated 2026-07-28, and the SDK page lists official kits in 10 programming languages, six of them in its top tier. Those kits matter to you because the MCP site says the standard "reduces development time and complexity" for the people building connectors, and every connector they ship is one more app your assistant can reach.

How is MCP different from an API, a plugin or a Zapier zap?

The difference is who decides the steps. An API or a zap runs steps someone fixed in advance. With MCP you ask in plain language, and the assistant picks which tool to call within the permissions you set.

	What it is	Who decides the steps	What you do to set it up	Example
API	The door one product opens for other software	The developer who writes code against it	Nothing; an automation tool or coding agent uses it for you	An n8n node reading rows from Google Sheets
Plugin	An add-on built for one host app	The host app	Install it from that app's menu	In ChatGPT today, "Plugins" is also the menu where you add an MCP server
Zapier zap	A fixed workflow: a trigger and the actions you drew	You, once, when you build it	Build it in Zapier's editor	New form entry, then a Slack message
MCP server	A standard adapter any MCP-capable assistant can use	The assistant, in conversation, with your approval	Toggle a connector or paste a URL, then log in	Claude finding a free slot in your Google Calendar

These overlap in practice. An MCP server is built on top of the product's own API, so the API doesn't go away. The MCP server makes it usable by any assistant. OpenAI's own help page has you add a custom MCP server and then "Create as a plugin", so in ChatGPT a plugin can now be an MCP server with a name and an icon.

Zapier sells both models. Zapier MCP lets "Claude, ChatGPT, Cursor, and any other AI client take real action" across Zapier's "9,000+ apps". I'd still use a zap for a job that has to run every Monday at 8 without me, and MCP when I start the job by asking. The three-step n8n or Zapier build shows the scheduled version.

What can you actually do with MCP today without coding?

You can let an assistant read and act in your email, calendar, files and notes from the chat window. The connectors below are documented on the vendor pages I opened, and none of them needs code.

Job	Connector	What the vendor page says it can do	Asks before acting?
Triage your inbox	Claude, Gmail connector	"Search and read emails", "Draft emails", send, reply and forward	Yes, for send, reply and forward by default
Book a meeting	Claude, Google Calendar connector	"Create, update, and delete events" and "Find mutual availability"	Not stated on the page
Find a file	ChatGPT, Google Drive app	Search files and reference Docs, Sheets and Slides	"ChatGPT may ask you to confirm" before changing a file
Update your notes	Notion MCP, hosted by Notion	Search, read, create and update content, create pages and databases	Not stated; set it in your assistant's tool settings
Reach a less common app	Zapier MCP	Actions across Zapier's 9,000+ apps	Set in Zapier and in your assistant

Sources for the table: Claude's Google Workspace page, ChatGPT's Google apps page, Notion's MCP docs and Zapier MCP.

A first useful request looks like this: "Find the three emails from last week where someone asked me for a call, check my calendar, and draft replies offering two free slots." The assistant searches Gmail, reads your calendar, and writes drafts. You read the drafts and press send.

In my own work, Google Ads and Meta run through MCP servers, so I ask for a performance pull in plain language and get a table back. Notion holds my pipeline and the assistant reads it before drafting a follow-up. I still read and send every message myself.

Which AI apps support MCP today?

Claude and ChatGPT are the two chat assistants here that you can set up as an individual, and Microsoft and Google document MCP in their business and developer tools. The table records only what each vendor's page said when I opened it on 15 October 2026, with the page's own last-updated date where it showed one.

App	What the vendor page says	Limits I saw	Page date
Claude (Anthropic)	Gmail, Calendar and Drive connectors "are available for all users on Claude and Claude Desktop". Custom connectors to any remote MCP server on Pro, Max, Team and Enterprise	Custom connectors are in beta. On Team and Enterprise only Owners add them	No calendar date shown
ChatGPT (OpenAI)	Add a server under Plugins, then "Add custom MCP server". Full MCP support on Business and Enterprise/Edu	Pro is limited to "read/fetch permissions". Web only. The page does not mention Plus or Free	No calendar date shown ("Updated: yesterday")
Microsoft Copilot Studio	Connect an agent to an existing MCP server	"Copilot Studio currently supports MCP tools and resources", and generative orchestration must be on	26 August 2026
Visual Studio Code with GitHub Copilot	Add, edit and manage MCP servers for Copilot	Warns that "Local MCP servers can run arbitrary code on your machine"	14 October 2026
Gemini Enterprise (Google Cloud)	An administrator sets up a custom MCP server as a data store	HTTPS only, and "Gemini Enterprise only supports the StreamableHTTP transport"	14 October 2026
Gemini CLI (Google)	Configure and use MCP servers in Google's command-line agent	A trust setting "bypasses all confirmation dialogs"	6 October 2026

Sources: Claude custom connectors, Claude Google Workspace, OpenAI help, Microsoft Learn, VS Code docs, Google Cloud, Gemini CLI docs.

I did not open a Google page on MCP in the consumer Gemini app for this check, so it is not in the table. If you have a personal account and no IT team, Claude is the easiest start: Claude's pricing page lists "Connect your apps and tools" on the Free plan, and the Google connectors need no paid tier. How the two assistants compare on everything else is in the ChatGPT and Claude everyday-work table.

What are the risks and what should you check before connecting a tool?

The main risk is that you give an assistant the power to act in your accounts based on text it reads, and some of that text can be written by a stranger. Anthropic and OpenAI both name prompt injection on their help pages: hidden instructions inside an email or document that try to make the assistant do something you did not ask for.

The MCP specification puts the duty on the app: "Hosts must obtain explicit user consent before invoking any tool." OpenAI's developer page says "ChatGPT currently requires manual confirmation in any conversation before write actions can be taken." Those approval prompts are where you catch a bad action, so read them before you click.

Before connecting anything, I run five checks.

1. Who built the server. Claude warns that custom connectors reach "services that have not been verified by Anthropic", and OpenAI says custom servers "are not developed or verified by OpenAI". Prefer a server run by the tool's own company, such as Notion's hosted one.
2. What the login screen asks for. Claude's help page says to review the permissions requested during sign-in and deny any that look unnecessary. A calendar helper has no reason to ask for your files.
3. Which tools write. Reading your inbox and sending from it are different risks. Switch off write tools you do not need in the connector settings, and choose "Allow always" only for tools you would trust unsupervised, as Claude's page advises.
4. Whether it runs on your computer. The VS Code docs warn that local servers "can run arbitrary code on your machine". As a non-coder, stick to hosted connectors you add by URL or toggle.
5. Whether you still use it. Claude notes that server developers can change tool behaviour unexpectedly, and the MCP remote-server guide says to "Regularly review and remove connectors you no longer use".

How do you connect your first MCP server?

Pick one app you open every day, connect it with read access only, and run one real task you can check by hand. In Claude none of the steps needs code.

1. Choose the job. One read-only task, such as "summarise every email from my accountant this month". Write down what a correct answer looks like before you start.
2. Turn on a built-in connector. In a Claude chat, click the plus sign, hover over "Connectors" and switch on Gmail, Calendar or Drive. Sign in with Google when asked.
3. Or add a custom connector. In Settings, open "Connectors", click "Add custom connector", paste the server URL from the tool's own documentation and click "Add". In ChatGPT the path is Plugins, then "Add custom MCP server", on the plans listed above.
4. Trim the permissions. Open the connector's settings and switch off every tool that sends, deletes or shares until you have used it for a week.
5. Run the task and check it. Compare the answer to the source yourself. If it missed an email or invented one, fix your request before you give it more rights.
6. Add one write action. After a week of clean reads, allow one write, such as drafting replies, with the approval prompt kept on.

The beginner module MCP: Plugging AI Into Your Actual Tools covers the metaphor, the first Gmail, Calendar, Drive and Notion connections, and how to tell tools that only read from tools that write, in 12 minutes. When you are ready to wire up one server end to end, Connecting One MCP Server (Consumer) walks through choosing the connector, the connection flow, how to give your first real task more access one step at a time and a permission audit. Both are part of the course in How do I learn AI without coding and get from chat user to operator?