MCP explained: how one server gives agents tools on 50+ platforms

One assistant, too many apps. Picture Lina, who runs a small online skincare shop. Her AI assistant already answers customers on WhatsApp and Telegram. Now she wants it to do real work: book a consultation in Google Calendar, send an invoice from Stripe, update the lead in HubSpot, and post daily sales to a sheet. Every one of those apps has its own API, the doorway programs use to talk to it, plus its own login and its own quirks. Wiring them one by one is slow and fragile. So what if there were one standard plug that let her agent reach all of them, and dozens more? That plug exists, and today you'll learn how it works.

What you will learn. By the end of this lesson you'll be able to do three things. First, explain what MCP is and why it saves you from building the same connection over and over. Second, describe the three parts of an MCP setup, the host, the client and the server, and follow a real request from your agent to a calendar and back. Third, plug one MCP server into a no-code agent so it can use tools on more than fifty platforms, and know which tools to switch on and which to leave off.

The many-to-many problem. Here's the core problem. Say Lina uses three agents and five apps. Without a shared standard, each agent needs its own link to each app, so three times five gives fifteen separate integrations to build and maintain. Change one app's API and several links break. With MCP, each agent learns the standard once, and each app gets one MCP server once. Now you add instead of multiply: three plus five, only eight pieces. And as you add apps, the saving grows fast, because somebody else has often already built the server you need.

MCP in one sentence. MCP stands for Model Context Protocol. A protocol is just an agreed format for messages, the way every email client agrees on how an email is shaped. MCP agrees on how an AI agent discovers tools and asks to use them. The best everyday analogy is USB-C. Your laptop doesn't care whether you plug in a charger, a screen or a drive, because they all speak the same port. MCP is that port for agents. Anthropic released it as an open standard, and today Claude, ChatGPT, n8n and many other tools speak it.

Host, client, server. Let's meet the three parts. The host is the app your agent lives in, like n8n or Claude Desktop, and it talks to the language model, or LLM. Inside the host sits an MCP client, a small connector. On the other side is the MCP server, which wraps real systems such as your calendar and your CRM. Now watch a request travel. First the client asks the server which tools it has, and the server answers with a list. Then the agent picks one, the client sends a call, the server books the slot, and the result flows back.

What a server offers. An MCP server can offer three kinds of things, so think of it as a well organised back room. First, the Tools. These are actions the agent can take, like booking an appointment, creating an invoice or adding a lead. Second, the Resources. These are things the agent can read, like your price list, a customer's order history or today's sales file. Third, the Prompts. These are ready-made instructions the server shares, like a template for a polite payment reminder. Most of what you'll use day to day is tools, but knowing all three helps you read any server's description.

Worked example: a call, step by step. Let's make this concrete. These messages are written in JSON, a simple text format for labelled data: a name in quotes, then its value. Part one is what Lina's server advertises: two tools, each with a name, a short plain-English description, and the fields it needs, like the date and the customer. A customer named Sara writes on WhatsApp asking for her invoice. The model reads the descriptions and picks create invoice. Part two is the message the MCP client sends: a tools call with that tool's name, and the arguments filled in, customer Sara and amount 120. Part three is the server's answer after it has done the work: Invoice for Sara, 120 dollars. The agent replies to Sara with exactly that.

One server, 50+ platforms. Now the title of this lesson. Some MCP servers wrap just one app. Others are gateways, one server that sits in front of hundreds of apps. Zapier's MCP server, Composio and Pipedream all work this way. You sign in to Gmail, Google Calendar, Stripe, HubSpot and Google Sheets once inside the gateway, choose the actions you want, and get one server address. Your agent connects to that single address and instantly sees tools on all those platforms. For Lina, that's booking, invoicing, the CRM and sales reports behind one plug, with no API code at all.

Plugging it into a no-code agent. Here's how it looks in n8n. Your WhatsApp or Telegram trigger feeds the AI Agent node, just like in the last lessons. Under the agent's tools, add the MCP Client Tool node, paste the gateway's server URL, and choose which tools to include. That's it. Servers come in two flavours. Remote ones you reach over the web by URL, like the gateway, and that is what n8n's MCP Client Tool connects to. Local ones run as a program on your own computer and are used by desktop hosts like Claude Desktop. And because the tools live on the server, your WhatsApp, Telegram and web channels can all share the exact same toolbox.

Common mistake: switching everything on. The most common mistake is switching on every tool a gateway offers. Two hundred tools means the model reads two hundred descriptions on every message. It gets slower, costs more, and starts picking the wrong one, like sending an email when it should have booked a slot. It's also a security risk: an agent that can issue refunds or delete contacts will eventually do it by accident. So enable only the five or six tools this agent truly needs, write clear descriptions, and put a human approval step in front of anything that moves money or deletes data.

Recap. Let's wrap up. First, MCP is a shared standard, so instead of building every agent times every app, you build every agent plus every app. Second, the host and its client ask questions and make calls, while the server lists its tools and actually runs them against your real systems. Third, a gateway server puts more than fifty platforms behind one address, which is powerful, so you choose only the tools each agent needs and guard the risky ones. With that, Lina's assistant can finally do real work, not just chat.

Your exercise. Here's your task before the next lesson. Pick one gateway, either Zapier MCP or Composio, and create a free account. Connect Google Calendar and one invoicing or CRM app you actually use. Enable exactly three tools, no more. Copy the server URL into an MCP Client Tool node attached to your agent from the earlier lessons. Then send it a test message like, book Sara for Friday at three pm. Check that the event really appears in your calendar. In the next lesson we'll bring everything together into a full omnichannel assistant that books, invoices and updates the CRM.