so what the heck are MCPs and skills?

Recently at work, I learnt about MCPs. Since last year, the term MCP has been blowing up everywhere, but I never really got the time to properly understand what it was until an opportunity finally arise at work.

At the start, I was confused. If MCPs are basically APIs for agents, why can't agents just call the API directly? And what the hell are skills? Aren't they just prompts stored in a .md file? Why can't I just store my prompt somewhere? And how is all of this different from something like a custom GPT?

After actually building with them, it started making a lot more sense.

The exact purpose of what I am building is confidential, but essentially there is a workflow that users always struggle with because it is tedious. There are multiple steps they have to remember, information they need to retrieve, and decisions they need to make along the way.

What I did was map that workflow into a skill, while whatever information or actions the AI needs are exposed through an MCP.

The easiest way I understand it now is:

Skill = how the work should be done
MCP = what the AI is capable of doing
Agent = decides what to do next

MCP isn't really replacing APIs. The API can still exist normally. The MCP is more like a standard layer that exposes useful parts of your system as tools that AI applications can understand and call. Technically, I could just write custom functions that call my API directly, but then every AI application would need its own integration. MCP gives these tools a common interface.

Skills confused me even more because, technically, they really can just be Markdown files containing instructions. But the useful part isn't the file format. A skill is basically taking the knowledge of how a task should be done and making it reusable.

Instead of repeatedly prompting an AI with things like "first check this, then retrieve this, if this happens do that, and make sure you verify this before finishing", I can define that workflow once as a skill.

This is where MCPs and skills started clicking for me. The skill contains the workflow and decision-making process, while the MCP provides the tools needed to actually carry it out.

For example, the skill might say to check whether something already exists, retrieve some information, make a decision based on it, and finally create something. The MCP might expose tools like search_existing, get_information, and create.

The skill knows what should happen, while the MCP knows how to interact with the actual system.

There is definitely some overlap with custom GPTs (from the ChatGPT app) too. A custom GPT is more like a configured assistant with its own instructions, knowledge and tools. A skill is a reusable workflow that an agent can follow, while an MCP exposes capabilities that potentially many different agents or skills can reuse.

I think the most interesting part for me now isn't really MCP itself. It is figuring out where things should live. What should be part of the skill? What should become an MCP tool? What should remain normal application logic?

The way I currently think about it is that if something describes how the agent should approach a task, it probably belongs in the skill. If something requires real data, system access, or a deterministic action, it probably belongs behind a tool.

After working with this, MCPs and skills feel a lot less like random AI buzzwords. They are really just ways of structuring agents so that the workflow, tools and reasoning don't all end up dumped into one giant prompt.