🧠 What’s the deal with MCP? And why you should start exploring it now

#mcp #ai #programming #typescript

A couple of weeks ago, I started hearing more and more about something called Model Context Protocol (MCP). Actually, the first thing I saw was a TikTok by ChicaPython, and from that I kind of understood that MCP was like an API… but for communicating with LLMs (Large Language Models) 🤯.

Then I started googling and found the official documentation, with several readings on the architecture of the protocol. From there, with some examples and tutorials using Claude, I jumped into the practical side.

And here’s what makes it powerful: MCP is a standard protocol, declared in November 2024 (so yeah, pretty recent). And being a standard means it’s here to stay. From now on, this is how we communicate with LLMs, not just to consume services like when we used the Gemini API to create an agent. We now use MCP to boost communication and increase model operability.

With MCP, you can create an agent and, through an MCP server, give it specific permissions to perform concrete actions. For example, you can authorize an agent to read and write files in a local folder—it can pull information from there and save outputs too. And the best part is, thanks to the Open Source community, there are already tons of MCP servers available:

To connect with GitHub and have the agent build a Pull Request for you (like Jules, the copilot presented by Google at this year’s I/O),
To read a database,
Or to automate tasks like checking issues, generating reports, and much more.

🟡 Side note: As with anything powerful, use it with good judgment. Review carefully what permissions you’re giving before using third-party MCPs. You might be exposing more than you think. (Let’s bring back the classic phrase: “With great power comes great responsibility.”)

💡 What I loved the most

You don’t need a powerful machine, weird programs, or to learn a brand-new language. None of that.

You can work locally, run the MCP inspector, and test what you’re building in TypeScript, for example or any compatible language (I’ll drop them at the end of the article). You can even configure an existing MCP with a simple JSON file. Easy-peasy.

You can also test the tools you develop directly integrated into Claude, for free.

That’s how I created my first MCP server in TypeScript, following a tutorial by Midu. I used a local viewer to test tools, activated developer view in Claude to test what I was building, automated GitHub pull requests using @modelcontextprotocol/server-github, and managed to create an agent connected to a weather API that gave me real-time weather for any city I asked about. I also accessed the local file system to read and save files in a controlled way.

The most exciting part? Just with the ability to read and write files, you can give your model context with information in directory X and save the synthesized results of your queries to directory Y.

All under control. No hallucinations. Clear steps.

All from your local machine!

Here’s an example of how the MCP config looks when enabling access to your file system, like giving it permission to read your Desktop. When you add this tool to Claude, you can interact with the agent, ask it to read files on the Desktop, give you outputs based on them, and save the results to files.

👀 So, why should you start exploring it too?

Not just because it’s here to stay, and not just for the hype. This goes deeper.

You should start exploring MCP because:

If you feel like you’re late: today is literally the earliest you can start.
You don’t need any Machine Learning background.
You don’t need to learn a new language: MCP has official SDKs in TypeScript, Python, Java, Kotlin, and C#—so chances are you can already start with what you know (and if not, maybe it’s the perfect excuse to learn something new through a fun project).
You’ll unlock creative thinking: playing with MCP will spark all kinds of ideas to connect tools and explore new possibilities.

Also: MCP is model-agnostic. You can use it with Claude, ChatGPT, Gemini, or whatever comes next.

The model itself isn’t the key. What matters is the context you provide, the instructions, the tools you expose, and the permissions you define.

🎮 Exploring is super easy

A single idea bouncing in your head is enough. You don’t need a huge project to get started.

You can wrap an API you already use, build a simple automated flow, or just play around with local files.

And if something feels uncomfortable along the way (like granting permissions, understanding what the model sees or can execute), that’s okay.

That discomfort teaches you too. Questioning boundaries is part of the journey and it’s how we shape more responsible AI.

🧭 100% recommended links to get started

Here are a few useful links if you want to begin (just like I did 👀):

MCP won’t do everything for you, but it gives you the power to decide how and what to build.

Have you tried it yet? If not, I hope this little blog post inspires you to start.

You could start by reading files from your local system, automating a GitHub pull request, connecting to a weather API, or even building a small tool to organize your notes.

It’s not about how big the project is, it’s about the curiosity to explore ✨