3.2 — Giving the AI "hands": tools

To act, an agent must be connected to your software. Two notions not to confuse:

🟢 In plain words — a tool is an action the AI knows how to trigger (search, calculate, send an email). MCP is not a tool: it’s the standard plug that lets you connect those tools without custom wiring. The tool = the hand; MCP = the socket.

• Tools (tools / function calling): the ability, given to the AI, to call an external function — a web search, a calculation, sending an email. The basic building block.
• MCP (Model Context Protocol): the universal standard for plugging in those tools and data sources, without recoding a custom integration every time. (Detailed in 3.3.)

🔧 Concretely, what is a “tool”? A function the AI can trigger by writing a structured call. Examples: rechercher_web("Paris weather"), envoyer_email(to, subject, body), lire_ligne_airtable(id). The AI doesn’t “click” like a human: it chooses which tool to call and with which parameters.

🧩 Concrete example — you ask “what’s the weather in Paris?” The AI doesn’t know it by heart. With a rechercher_web(...) tool, it writes the call rechercher_web("Paris weather"), receives the result, then answers you. Without that tool, it could only guess or admit it doesn’t know.

🧪 Mini-exercise (1 min): list 3 “tools” YOUR project would need to act (e.g. read an inbox, look up a price, write to a spreadsheet). It’s the list of “hands” of your future agent.

❓ Check your understanding — are a tool and MCP the same thing?

See the answer

No. A tool is a precise action the AI can call (search, send, read). MCP is the standard used to plug in those tools in a uniform way. You can have tools without MCP; MCP just makes the connection universal and reusable.

✅ In short

• A tool = an external function the AI triggers (search, email, calc…).
• The AI chooses the tool and its parameters; it doesn’t “click” like a human.
• MCP is the standard plug to connect those tools (detailed in 3.3).