MCP in Claude: connect Telegram, Sheets, and your site to AI

Section 01What MCP is, in plain words

MCP stands for Model Context Protocol. Behind the scary name hides a simple thing: it's the standard by which AI connects to outside apps. Anthropic – the company that built Claude and invented MCP – explains it like this: it's a USB port for AI. Every app used to have its own tricky way to connect, and hooking it up to AI took a programmer. Now there's one shared plug that anything slots into.

To make it fully clear, compare it to a phone. A bare phone can call and text. But the moment you add apps – bank, maps, messenger, notes – it becomes a remote for your whole life. MCP does exactly that to AI. Bare Claude can think and write text. Connect your services through MCP and it becomes a remote for your work.

Who's talking to whom

MCP always has two sides. On one is the AI – for Anthropic, that's Claude. On the other is a small translator program called an MCP server, or connector. The connector knows how a specific service is built: how to pull a row from Google Sheets, how to send a message in Telegram, how to read an email in Gmail. Claude tells the connector what it needs in a shared language; the connector translates that into the service's language and does it.

You don't need to know how the connector works inside. For ready services these connectors are already written, and your job is to switch them on with a checkbox. It's like a video call: you hit the green button, and which protocols carry the picture and sound is none of your concern.

Why this matters right now

A year ago, for AI to do anything inside your services, you had to order an integration from a developer. Expensive, slow, and it broke on any change. MCP took the developer out of that chain for everyday tasks. Services now ship their own connectors, and you hook them up in a couple of clicks. That's exactly why there's so much noise around MCP: it opened up to an ordinary person what used to cost money and require code.

Here's a live example from my own work. Every morning I needed one number: how many leads came in yesterday and how many of them paid. I used to open the spreadsheet, count by eye, ballpark the conversion. Five minutes, but every day, and often wrong when I was half-awake. Now Claude checks that same spreadsheet through a connector and by seven a.m. drops three lines into my Telegram: leads, payments, what to watch today. I don't open that sheet for weeks. That's MCP in action: the spreadsheet is the same one, but the data now walks to the AI on its own, with no me in the middle.

Section 02What changes when AI has hands

The difference between a normal chat and Claude with MCP is like the difference between an advisor and an assistant. An advisor tells you what to do; you still do it. An assistant just does it. Let's take three everyday tasks so the difference gets tangible.

Before and after

• The sales report. Before: you open the spreadsheet, copy the rows, paste into the chat, get the breakdown, move the conclusion into your notes. After: you send Claude one sentence, it checks the spreadsheet itself, counts the numbers, and sends the takeaway. You never even open the sheet.
• A reply in DMs. Before: a client asks a question, you read it, go to the chat for the wording, copy the answer back into Telegram. After: Claude sees the message through the connector and answers the client itself, in your tone.
• An edit on the site. Before: you ask AI to rewrite the text on a page, copy it, dig into the site's admin panel, paste, save. After: you tell Claude to change the text, it opens the page and edits it itself.

Notice what all three left-hand columns share. In every one, you are the courier between the AI and the service. It's exactly that courier work MCP takes off you. AI stops being a picture in a window and starts reaching into the real world of your apps.

Two actions, not one

The key thing to grasp: MCP gives two abilities at once. First – reading. Claude can look into your spreadsheet, read an incoming email, see what a client wrote. Second – acting. It can write a row, send a message, update a page. Weak integrations usually only read. The power of MCP is that AI closes the loop: it saw, it understood, it did. From that grows everything else. A combo where AI both reads and acts is already almost an AI agent. It's only missing the autopilot – doing it on a schedule or on an event, without your command. How to take a combo all the way to full autopilot is covered in how to build an AI agent with no code.

Section 03What you can connect: the connector catalog

There are already hundreds of connectors, and more every month, because MCP is open and services build the connections themselves. Let me lay out the most useful ones for someone who sells services, courses, or consulting, grouped by type. Don't try to connect everything at once – the one-connector rule comes up later in the article.

Three kinds of connectors

Connectors come in three kinds, and the kind decides how simple the setup is.

1. Official, built-in. Built by the services themselves or by Anthropic. Google Drive, Gmail, and calendar usually already sit in Claude's settings, switch on with a checkbox, and only ask you to sign in. The simplest kind.
2. From the catalog, ready-made. Written by the community and posted in open connector catalogs. Telegram, databases, and various niche services usually live here. They install as an extension and sometimes ask you to paste in a key or token from the service.
3. Your own, for your task. If there's no ready connector for your unique site, Claude Code builds it – that's the version of Claude that can write and run code for you. You explain in words what you need; it writes the code. What this tool is and how to start is covered in Claude Code for non-coders.

The rule is simple: first look for a ready connector from the first two kinds, and only if none exists do you go to the third, to Claude Code. Nine times out of ten, a ready one turns up.

Section 04How to connect with no code: three ways

Now, step by step, how it's done by hand. The ways run from simplest to most powerful. Start with the first; you'll get to the third when you want to connect something of your own.

How long the first connection takes depends on the way. A ready Google connector is three minutes: found it, clicked it, signed in. Telegram through a token is about fifteen minutes, including setting up the bot. Your own connector to a site through Claude Code is more like an evening, because Claude first figures out how your particular site is built. But it's one-time work: connect it once and the combo lives on its own, taking no more of your time.

Notice that even in the hardest, third way you don't write code. You talk to the AI in words and forward it credentials. It's the same principle as in every article on this blog about building assistants: the AI handles all the technical part; your part is to explain the task and confirm access.

Section 053 MCP combos for your work

Let's take three combos that pay off first. Each is a connector plus a task. Pick one, build it, live with it a week, then add the next.

Combo 1. Telegram: AI answers your DMs itself

The most lucrative combo, because it's about sales. The same questions land in your DMs: how much, how to join, what's included, is there a payment plan. Through the Telegram connector, Claude sees those messages and answers them itself, in your tone, from your playbook. The moment someone's ready to talk or asks a question outside the playbook, the bot calls you in.

It's built through a bot and a token, like Way 2 above. You hand the AI a base of answers: prices, terms, common questions. From there it works off that base around the clock. How to take an assistant like this all the way from start to finish is covered in the AI agent in your Telegram DMs.

What it gives you: DMs answer themselves, and you step in only where the deal is decided. One evening of setup takes hundreds of identical replies a month off your plate.

Combo 2. Sheets: a sales report without opening the spreadsheet

Everyone who sells has a spreadsheet with leads, payments, and ad spend. The data sits there, but the conclusions don't, because there's never time to look. Through the Google Sheets connector, Claude checks the sheet itself, counts the key numbers, and sends you a human takeaway.

It connects through the official Google connector – sign in and you're done. From there you just ask how sales are going, and Claude opens the spreadsheet and answers. On autopilot you can set it so the report arrives every morning by itself. How to read the numbers so they change your decisions is covered in the models for making money with AI.

What it gives you: you start seeing your numbers instead of fearing them. And it takes a minute, not an evening of reconciling rows.

Combo 3. Site: AI writes and updates pages

A blog and a site need constant edits: add an article, change the text on a page, update a price. Usually that means digging into the admin panel and doing it all by hand. Through a connector to the site, Claude does it itself: you say what to change, it opens the right page and edits it.

This is a third-kind combo, built by Claude Code for your specific site. Set up the access once and from there edits go by voice or text. How this tool is built and where to start is covered in Claude Code for non-coders, and how to build a whole site by talking to AI is in building a site by voice.

What it gives you: the site stops being a place you're scared to touch. Edits that piled up for weeks get done in a five-minute conversation.

Section 06Security: what not to hand over

Once AI gets hands, it has access to your data and the ability to change it. That's not a reason to fear MCP; it's a reason to connect it with your head on. Keep four rules and everything stays calm.

• Least access. Give the connector exactly what the task needs, no more. If Claude only needs to read a spreadsheet, don't let it delete. Permissions are granted at connection, and you can always see there what you're allowing.
• Confirm actions. At first, keep the mode where Claude asks permission before every action that goes out: before sending an email, before an edit on the site. Once you're sure it doesn't slip, you can loosen it.
• No sensitive data. IDs, full client lists with phone numbers, trade secrets – don't run those through connectors. Anonymize: swap real names and numbers for labels.
• Check what goes out. Anything headed to clients or to money, eyeball yourself at first. AI sometimes states falsehoods with full confidence, and on the important stuff it's better to keep your own eye on it.

A separate word on tokens and keys. That bot token or service key is the key to your apartment. Don't forward it into open chats, don't post it in screenshots. If a key leaks, you can always revoke it and issue a new one. What you should never load into AI, ever, is covered in detail in protecting yourself from AI-driven hacks.

Section 075 connection mistakes

Almost everyone starting with MCP steps on these rakes. Knowing them ahead of time saves you a week of nerves.

Mistake 1. Connecting everything at once

The temptation is big: if you can connect ten services, connect ten at once. That gets you a mess Claude gets lost in and you can't tell what it pulled from where. Connect one connector at a time. Once you've got it and checked it works right, add the next.

Mistake 2. Granting full access on day one

When the connection asks you to pick permissions, it's easy to hit allow all just to skip the thinking. Do the opposite: start with the minimum. You can always widen access later, but rolling back an extra permission is usually remembered only after something got deleted.

Mistake 3. Expecting a miracle from the connector

A connector gives AI hands, but the head is the same one. If you haven't explained your rules and given Claude context, it'll act blind. The strength of a combo isn't in the connection. It's in how well the AI knows your voice and your tasks. How to set up that head through a knowledge base is covered in Claude Projects.

Mistake 4. Sending things out without checking

Letting Claude send emails and post right away, with no glance from you, means catching a blooper in front of a client one day. For the first few weeks, keep confirmation on before every action that goes out. Trust is earned through work, not one lucky answer.

Mistake 5. Building your own when a ready one exists

Sometimes you want to dive straight into Claude Code and build your own connector. Usually that's wasted effort. First check whether a ready one exists in settings or the catalog. Building your own only makes sense for what genuinely isn't there – your unique site or a rare database.

Section 08Connection checklist

Save this and run through the points as you connect a single connector. When all seven boxes are ticked, the combo works and takes real routine off you.

• Task named in words – you know exactly what the AI should do and in which service
• Connector chosen – found a ready one in settings or the catalog; go to your own site only if none exists
• Permissions granted at the minimum – only what the task needs, no more
• Key or token in place – if the service needs one, you got it and pasted the string in
• Context gathered – the AI knows your voice, prices, rules through a Project
• Action confirmation on – Claude asks permission before anything goes out
• Combo run in for a week – you've confirmed it really saves time, and only then loosened control

How this connects to the other articles

MCP is the hands of AI. The head is shared: the better Claude knows your voice and rules, the cleaner every connector runs. How to set up that head is covered in Claude Projects. When you want to take a combo to full autopilot so it runs on its own, start with the guide on how to build an AI agent with no code. And the wider view of which AI-and-service combos take routine off you is gathered in automating your routine with AI. Three articles, one system.