MCP: give AI direct access to your real tools

This article is part of the Devlights Tech Tips series on AI applied to technical work. If you missed the previous episodes: model, agent and tool · tokens and context window · AGENTS.md · Skills · llms.txt.

📺 Watch the video on Instagram

Up to now, AI worked with whatever you gave it. MCP gives it direct access to your real tools, so you're no longer the go-between for the AI and the rest of your stack.

What is MCP?

MCP stands for Model Context Protocol. It's an open standard created by Anthropic that defines how to connect an AI model to external tools.

The key word is standard. It's not a proprietary integration that only works with one specific model or tool. It's a protocol that any tool can implement and any model can consume.

It works like an API: each tool you want to connect (GitHub, Notion, Slack, Drive) has its own MCP server. That server is the bridge between the model and the tool.

Model  →  MCP server  →  Tool
              ↑
        adapter that translates
        the model's actions into
        the tool's language

You configure which servers the agent has available. Then it decides when and how to use them depending on the task.

How it works under the hood

Each tool you want to connect has its own MCP server, basically an adapter that translates the model's actions into that tool's language.

When the model needs to do something (read an issue, open a PR, send a message), it talks to the corresponding MCP server. The server runs the action and returns the result to the model.

Without MCP:  model → you → tool → you → model
With MCP:     model → MCP server → tool → model

The difference is structural: with MCP, you stop being the middleman. The agent goes straight to the source, runs the action and keeps working, with no interruptions, no copy-pasting and no context switching.

In practice: Claude + Notion

To understand the real impact, nothing beats seeing it in a concrete case.

Say you have Claude Code connected to the Notion MCP server, where the team tracks the day's tasks. Instead of opening Notion, checking the board, copying the tasks and pasting them into the model, you just tell it:

"Go into Notion, check today's tasks, identify the highest-priority ones and give me a summary of where to start."

What the agent does:

1. Accesses Notion through the configured MCP server.

2. Reads the actual tasks on the board: not a copy, the live data.

3. Analyzes priorities based on the context.

4. Returns an actionable summary with the recommended order for tackling the day.

Notion was never opened. Nothing was copied. There was no middleman. The agent went straight to the source, processed the information and brought back the result.

The MCP ecosystem today

Notion is just one example. The MCP server ecosystem already covers most of the tools tech teams use:

And most importantly: you can build your own MCP servers to connect AI to your company's internal tools: the ones with no public integration, legacy systems, the business's own APIs.

That's what sets MCP apart from traditional integrations: you don't depend on the tool vendor building the connector. You can build it yourself.

Wrapping up the foundations

With this episode, we wrap up the series' six foundational concepts:

Now you know how each piece works. The most interesting part comes next: how each team uses all of this in its real workflow.