Your first MCP server, your first Skills, your first multi-agent system

Your AI can’t open a file

In 13 hours you go from copying and pasting into a chat window to having 5 MCP servers, your first Skills and a multi-agent system running on your machine. In Python.

Language: the course is taught in Spanish. The videos, the cheatsheets and the exam are in Spanish. This page is the English description only.

MCP is the protocol Anthropic published in late 2024 so a model can read your repository, query your database and call your APIs without you pasting anything into the chat. Cursor, Zed, Sourcegraph and half the tooling industry already speak it.

Everyone mentions it on LinkedIn and almost nobody has written one. That’s where this course gives you an edge.

It’s for you if you program in Python at an intermediate level and use a model as a user. You don’t need to know anything about MCP or agents: you start from zero and finish with things of your own working.

It isn’t for you if you’ve never written code, because I assume you can write a function and call an API. Nor if you want to build a chatbot with the OpenAI API, which this isn’t about. And if you need MCP at bank scale, with OAuth and Kubernetes in front, I haven’t recorded that yet.

I want it to read my repo by itself

Monday morning

You open Claude Code, ask it something about the repo at work, and the model queries the database, finds the function through the syntax tree and answers with real data. You haven’t pasted anything.

That’s what’s on the other side of these 13 hours.

You end up with five MCP servers written by you, in Python and with their tests. The first talks to a real database. The second wraps a REST API with its authentication. The third indexes your own code by AST, so the model understands your repository without you pasting files. The fourth edits firewall rules on Linux, Windows and macOS from natural language, with a dry-run mode so you don’t get a nasty surprise. The fifth reorganizes your disk, detecting duplicates by content hash, not by name.

You also get your first Skills for Claude Code, written so they activate on their own when they should instead of ending up forgotten in a .md nobody reads.

And an orchestrator with three agents working at the same time on the same task, with the patterns that hold up and the ones that always break.

All of it runs, is tested and you can show it on Monday at work.

I want to show it on Monday

What you’ll save yourself

You’ve had Claude or ChatGPT open in a tab for months. It helps, quite a bit at times, but it touches nothing of yours: you open the file, paste it, read the answer, apply the change, copy the error and paste it again.

The model thinks and you do the dirty work.

And the day you get fed up and decide to automate that trip, you find that connecting tools to an OpenAI model isn’t done the same way as at Anthropic, and Gemini goes its own way. You change provider and rewrite half your code.

Until recently there was no standard way to do this, so it isn’t your fault. Now there is one, and what you save here are the fifteen detours that cost me a month before I found the one that holds.

Who is selling you this

My name is Daniel Alfocea. With Alfonso Muñoz I wrote MCP Seguro, and I have more than a hundred tools published on GitHub, one of which ended up inside Kali.

That’s the résumé, and it only serves one purpose: so you know what’s inside wasn’t copied from the documentation. I build MCP servers for people who pay, they break on me, and what you’ll see is what was left standing after breaking them.

What isn’t in the official documentation

Almost everything else is. These things aren’t, because they come from fighting with them:

Why an MCP server that works perfectly on your machine falls over as soon as you connect it to someone else’s client, and which part of the configuration is almost always to blame.

The difference between a Skill and an MCP Tool, which sounds like a nitpick and is the reason many people build the second when they needed the first.

When MCP isn’t the answer and what you need is a plain old CLI, even if the one using it is an AI.

What tool poisoning is, how someone slips instructions inside the description of a third-party tool, and the five questions I ask before installing an MCP server I didn’t write.

Why the agent that steps on itself is the most expensive anti-pattern in multi-agent systems, and how to see it coming before it eats your token budget.

I won’t give away any answer here. They’re all inside.

What it costs

€180, once.

You bill that in a morning, or two if you’re stretched. It pays for itself the first afternoon you don’t spend fighting a handshake you don’t understand.

For that money you get 13 hours of video in 10 modules and 4 exercises with their statement and answer. It also comes with 3 PDF cheatsheets, more than 25 code examples you can copy and that work, and a final exam with a verifiable certificate.

You also join the platform’s community, where questions that don’t fit in a video get answered. And you write to me if you get stuck.

Access is for life. When I update the course, you get the new version without paying again.

I'll pay the €180 and join

Daniel Alfocea

P.S. If what holds you back is the price, do a quick calculation: how many hours have you spent this year copying and pasting code into a chat window. Multiply them by what you earn per hour. €180 is what it costs you to stop doing it.

I don't want to lose another month like this