Without MCP, the AI can only chat. With MCP, it acts: reads emails, queries a database, searches a web page, sends a Slack message. The trick is that you don’t connect one action at a time—you connect a server (an entire service, such as email) and it already announces all its tools at once. The agent sees the menu and chooses the right dish for the request.

Illustrative diagram — one connection (the server) provides a set of tools. The agent reads what’s available and decides which tool to use for each request.

What it is

MCP (Model Context Protocol) is a common language for AI to communicate with external services. Instead of wiring up one action at a time, a single connection exposes many tools from that service. Think of an email service: when you connect the email server, the agent gets access to "send," "draft," and "search messages"—all at once.

The protocol solves the problem of “how AI discovers and uses an action”: it standardizes tool descriptions. That’s how the agent can decide on its own which tool to call, when call and with which parameters. This is how AI gets "arms"—it goes from just chatting to actually acting in the world.

Why learn

Key concepts

What it is

As soon as the agent and MCP server connect, the agent begins to know all available tools from that server — names, descriptions, and parameters. From there, the decision is up to it: you describe the result in natural language ("send the team a notification"), and the agent chooses the right tool ("send message") and fills in the parameters.

Illustrative diagram — you don’t say “use tool X.” You describe the result; the agent compares the request with its known tools and chooses the best one.

Why learn

Knowing the agent makes the decision changes how you ask: you focus on the result, not in pointing out the tool. And when it chooses incorrectly, you correct it with context ("don't draft it, actually send it") instead of micromanaging. This also explains why a server that's "on but useless" still gets in the way: your tools are added to the list the agent has to evaluate.

Key concepts

What it is

Installing an MCP server creates (or edits) a file .mcp.json: it stores the server configuration and, when the service requires authentication, a space reserved for the API key. You paste the key itself manually in the file. Never in chat, and never ask the agent to enter it for you.

{
  "mcpServers": {
    "firecrawl": {
      "command": "npx",
      "args": ["-y", "firecrawl-mcp"],
      "env": {
        "API_KEY": "<cole-sua-chave-aqui>"
      }
    }
  }
}

Why learn

A key pasted into chat becomes part of the conversation history—which may be logged, shared, or leaked. Pasting it into the file keeps the secret where only you (and the program) can access it. This is a security rule, not a style preference: treat every key like a password.

Key concepts

What it is

MCP configuration can live in two scopes: global (applies to the whole machine; all projects can see it) or by project (.mcp.json stays inside the project folder and only applies there). The recommendation is to by project: it isolates the secrets and context for each job.

Illustrative diagram — globally, one leak contaminates everything; per project, each job stays in its own bubble.

Why learn

Global configuration can trigger a security warning on every restart (the program warns that global servers are active for every project). The fix is to move the config into the project folder. Project by project, you also prevent one client from “seeing” another client’s tools and keys—less risk, less noise.

Key concepts

What it is

MCP servers are found in two main places: MCP directories (sites that catalog servers, like mcp.so) and in the GitHub (search for "MCP server" + the name of the service you want). Before installing, look for the target service's server; then follow its README.

Why learn

Knowing where to look saves you from reinventing the wheel: in most cases, the server you need already exists. Knowing the common categories also speeds up the search—you know there's something for web scraping, databases, documents, and communication, so you go straight to it.

Common server categories

Key concepts

What it is

The command /mcp lists the installed servers and the status for each one (connected, with an error, etc.). That’s your "MCP health." And there’s a golden rule for diagnosis: if a newly installed server doesn’t appear in the list, restart the assistant first — that solves it most of the time.

/mcp

The complete sequence, step by step

Why learn

Without the /mcp, you’re left in the dark: you don’t know whether the server connected, failed, or wasn’t even read. And without the habit of restarting, you waste time “fixing” what isn’t broken. Restarting is the first step for any MCP troubleshooting—cheap, fast, and almost always enough.

Key concepts

What it is

Each server active injects your tool definitions into the context from AI — and that consumes tokens. More connected servers mean more text the model has to read before responding, even if you don’t use any of them. The rule: turn off the servers you're not using to free up context.

Illustrative diagram — each server connects a block of definitions. Fewer blocks mean more room for the AI to think, lower costs, and less noise.

Why learn

Cluttered context has three costs: it uses more tokens (more expensive), confuses the agent (more tools to evaluate means a greater chance of choosing the wrong one), and slows everything down. Keeping only the servers needed for the current task active is the simplest way to reduce cost and errors at once.

Key concepts