What a Tool Actually Is
In plain terms, why AI needs tools, and the actual path a request takes from you, to the agent, to the AI, and back again.
AI can only talk — it can't do things
An AI model like ChatGPT or Claude does one thing: it reads text and writes text back. That's the whole job. It can't check today's weather, do exact math, or send an email, no matter how you phrase the question. If you ask it something it genuinely doesn't know, it doesn't say "I don't know", it often just guesses a plausible-sounding answer instead.
A tool is how that gets fixed. It's a small helper, a search, a calculator, a calendar app, that can actually go do the real thing and bring back a real answer. The AI still can't press the button itself. Something else has to do that on its behalf.
Who's the "something else"? The agent
That in-between piece is usually called the agent. Think of it as an assistant standing between you and the AI, watching the conversation and doing the actual legwork whenever the AI needs something done. You never talk to it directly, but it's there on every request.
The full round trip, step by step
Say you ask, "What's the weather in Paris?" Here's everywhere that question actually goes before you get an answer:
YOU AGENT LLM
│ │ │
│ "What's the weather │ │
│ in Paris?" │ │
├───────────────────────>│ │
│ │ Passes along your │
│ │ question, plus the │
│ │ list of tools it's │
│ │ allowed to use │
│ ├───────────────────────────>│
│ │ │
│ │ "Call the weather tool │
│ │ for Paris" │
│ │<───────────────────────────┤
│ │ │
│ │ Actually runs the │
│ │ weather tool, gets a │
│ │ real answer: 18°C │
│ │ │
│ │ Sends that real answer │
│ │ back to the AI │
│ ├───────────────────────────>│
│ │ │
│ │ "It's 18°C and cloudy │
│ │ in Paris right now" │
│ │<───────────────────────────┤
│ "It's 18°C and cloudy │ │
│ in Paris right now" │ │
│<───────────────────────┤ │
Notice the AI (the LLM) never leaves its box in the middle. It only ever reads text and writes text. Every "action" you see, actually calling the weather tool, actually getting a real number back, happens in the agent, one step to the side. To you, it feels instant and direct. Underneath, your question bounced through two extra stops before you saw a reply.
This is the pattern behind basically every AI tool you've used: a chatbot with web search, a coding assistant that runs your code, an app that checks your calendar for you. Different tools, same three-stop loop.
How does the AI even know a tool exists?
The agent tells it, in plain text, before the conversation starts. Something like:
Tool: weather_tool
What it does: Gets the current weather for a city.
Needs: a city name
Gives back: temperature and conditions
That's really it, a name, a one-line description, what it needs, what it hands back. The AI reads that the same way it reads everything else: as words. If the description is vague or the inputs aren't clear, the AI is more likely to use the tool wrong, so a good description matters more than it might sound like it should.
Next time you ask an AI assistant something that needs a search or a calculation, picture the diagram above. See if you can guess, from the answer you get, whether it actually called a tool or just guessed.
Where MCP comes in
So far, none of this needed MCP, it's just how tools work in general. The catch shows up once there's more than one AI assistant, or more than one tool: each pairing needs its own version of that "here's what I can do" description, built by hand, every time. MCP is what makes that step reusable instead of rebuilding it for every pair. The next chapter picks up exactly there.