The Chatty AI 🗣️🤖

Back in 2021, when we first started playing around with AI, what we really had was basically a know-it-all bar buddy. Sure, a very cultured, very technical one—capable of explaining everything from how to spin up a server on AWS to how to make a Spanish omelet with onions (mistake).

The hype was so big that nobody noticed what quickly became obvious. It talked a lot, but did very little.

AI was (and still is, to a large extent) like that friend who always has the perfect theory… but in the end, you’re the one stuck with the dirty work.

👉 You: “How do I configure my LAMP server and install WordPress?”
👉 AI: “Easy! Just run these 25 commands, edit 14 config files, and in about 3 hours you’re done.”

Spoiler: you still end up sweating in front of the terminal while it just keeps smiling.

And that’s when a new vision started to pop up:
“Okay, but couldn’t AI just do it for me? Can’t it already buy me concert tickets, book my table, or fix the damn bug?”

The answer today is still: yes… but not really.

Because “AI” does a lot of things these days—but careful: it’s not really AI doing them. It’s not ChatGPT, Gemini, or Claude… or did you think it was? Keep reading and I’ll explain 👇

The Age of Agents (But Without the Magic 🧙‍♂️)

Then came the so-called “agents.”

The idea sounded great: “AI doesn’t just tell you what to do, it actually does it for you.”

But here’s the trick: agents are really just plain old programs that interpret what the model says and then execute it.

Yep, you heard that right… an agent “talks” to the AI, then interprets its response and runs the actions needed.

Simple example:

👉 You ask: “Where do I have the file invoices.xlsx saved?”
👉 The agent asks the AI the same question, with a bit of tuning. (See below.)
👉 The AI gives the agent a command to run.
👉 And it’s the agent who actually executes the command, grabs the result, and passes it back to the AI.

¿Magia? Un poco, pero en realidad se puede hacer mejor. Mucho mejor.

Magic? A little—but we can do better. Way better.

AI still isn’t “doing anything.”

The one doing the work is the agent—playing personal assistant, but really just a script on steroids.

And while this has enabled tons of flashy demos, underneath it’s still duct tape: every agent invents its own way to wire AI to the real world.

Until… MCP shows up. 👇

The Real Leap: MCP 🚀

And here’s where it gets interesting.

After so many “yes, but not really,” something shows up that might actually change the game: MCP (Model Context Protocol).

It was invented by Anthropic (the folks behind Claude, their AI), and the idea is simple but powerful: create a standard protocol so AIs can connect directly and securely to services, apps, and tools.

👉 Now those “scripts on steroids” we talked about? Gone. The agent is no longer just a script. The agent is the AI itself. ChatGPT, Gemini, Claude… or even an AI running on your own laptop or phone.
👉 In other words: with MCP, you’re giving AI the keys to actually do things, not just talk about them.

Back to the earlier example:

Before, AI would say, “Run this command in your terminal,” and the agent would do it for you.
With MCP, you can grant direct access so the AI itself runs those commands—in a controlled, secure way, without making up weird stuff.

In short: MCP is the official bridge between AI and the outside world.
The path that lets it step out of the chat and into digital reality.

Think I’m overhyping it? 😏
Keep going—this was just a warm-up. The good stuff comes next.

A Truly Disruptive Example with MCP: an Agent that Buys Online 🛒⚡

Ever wondered why, even today, AI still can’t just buy you those concert tickets you want?

Or better yet: why, even though there are already “shopping agents” out there, this hasn’t really taken off? (If you’ve tried them, you know—reliable is not the word you’d use).

The reason is simple: the current process is a Frankenstein.

👉 The agent opens the website and grabs its contents (text, images, buttons).
👉 The AI interprets what it sees and replies: “Click the Buy button.”
👉 The agent clicks, re-captures the screen, and sends it back to the AI.
👉 The AI interprets again and says: “Ask the user for their name and fill in field X.”
👉 And so on… an endless loop of action → interpret → action → interpret.

Is it possible? 👉 Yes.
Is it efficient? 👉 Less than walking to your hometown on foot.

And it’s a real headache, because every website is different, and interpretations are bound to go wrong.

Classic example:
You’re on a product page and the AI see two “Buy” buttons:

— one for the actual product,
— and another for the ridiculous upsell they’re trying to sneak in.

Result: the AI gets confused and… surprise! Instead of your SSD, you end up with a Bluetooth blender.

So, How Do We Build a Real Shopping Agent? 🛍️🛠️

Here’s where the magic comes in: the solution to the chaos is called MCP.

👉 Imagine that instead of screenshotting pages and interpreting text like a tourist trying to read a menu in Mandarin, the AI actually knew exactly how shopping works in every online store.
👉 Imagine it didn’t depend on the browser or on blindly clicking buttons.
👉 Imagine it had an official, dedicated channel to talk to the store and say: “Hey, I want to buy this product for this person.”

That channel is MCP.

Anyone who’s ever shopped online knows the whole process boils down to just a few basic actions:

• Add to cart
• Enter shipping details
• Enter payment details
• Confirm purchase

With MCP, you define these actions clearly and in a standardized way, so the AI understands them and can execute them—no guessing, no interpretation.

👉 It no longer has to figure out which button is the right one.
👉 It doesn’t care if the HTML changes or if someone pops up an annoying “Subscribe to our newsletter.”
👉 It just uses that special channel and performs the action directly.

The result:
No guessing.
No ambiguity.
👉 Finally, a shopping agent that’s fast and reliable.

That clean, well-defined set of actions with zero room for error? That’s what we call MCP.

Still Not Convinced? 🤨

Alright, let’s keep going…

Those four basic actions, that MCP, that special channel that lets AI talk directly to the store… now picture this:

👉 Shopify implements it—so it’s available across all its stores.
👉 WordPress/WooCommerce does the same.
👉 Magento jumps on board.
👉 Even Prestashop joins the party.

See it now?

Overnight, because MCP works the same way everywhere, AI can buy for you in 99% of online stores.

Starting to sound a bit more impressive now, isn’t it? 😉

But Isn’t This Just an Old-School API?

Good question. And yes—if we’re only talking about online stores, it does look a lot like a regular API.
The difference is that MCP isn’t limited to “online stuff.” It’s a standard designed for any interaction AI might have with your systems.

👉 Imagine giving it an MCP that connects to your database (schema included) → now the AI doesn’t just answer questions, it generates business reports on the fly.
👉 Imagine another MCP wired to your ERP → it can issue invoices, register orders, or even balance your books for you.
👉 Imagine yet another MCP connected to your server’s command line → suddenly you’ve got a sysadmin working 24/7, who never sleeps, never complains, and never asks for a raise.

And the same goes for absolutely anything else you can think of.

No Way… This Won’t Happen in a Hundred Years…

You think so? 🤔

Alright, here are some facts:

👉 MCP was created by Anthropic, the folks behind Claude.
👉 OpenAI jumped on board—you can already use it in their custom GPTs.
👉 Gemini has started integrating it too.
👉 Shopify has already rolled it out on its platform.

So no, this isn’t science fiction.
This is happening right now—and pretty soon, you’ll be seeing it everywhere.