Tutorials · Chapter D (4/4) · ~11 min

Tools in code

Play → read → next

Wire a tool call in Python — model chooses, your code runs, result returns.

Code Lab

Tools in code

Wire function calling: tool name → function → result → reply.

Python

# Tools in code — model chooses a tool name + args
def weather(city):
    return f"Weather in {city}: 28C sunny"

def calculator(expr):
    # tiny safe demo calculator for a few expressions
    allowed = {
        "17*24": "408",
        "17×24": "408",
        "10+5": "15",
        "100/4": "25",
    }
    return allowed.get(expr.replace(" ", ""), "0")

TOOLS = {
  "weather": weather,
  "calculator": calculator,
}

def run_tool(name, arg):
    return TOOLS[name](arg)

# Pretend the model requested a tool:
call = {"name": "calculator", "arg": "17*24"}
result = run_tool(call["name"], call["arg"])
print("tool result:", result)
print("final reply: The answer is", result)

# Challenge: change the call to weather for Dubai

Recap

What you just did

You implemented the tool loop builders ship: schemas the model can choose, a dispatcher in your code, and a second model turn with the tool result. That’s how agents stop hallucinating the weather and start reading an API.

Teach

How it works

def get_weather(city: str) -> str:
    return f"72F and clear in {city}"  # stand-in for an API

tools = {"get_weather": get_weather}

# model might return: name="get_weather", args={"city": "Lisbon"}
name, args = "get_weather", {"city": "Lisbon"}
result = tools[name](**args)
# send result back in the next messages turn
print(result)

Declare tools (name, args, description)
Model selects one (or none)
Your code runs it — never trust invented results
Model answers using the real output

Mental model: LLM proposes; Python disposes.

Use it

When you'd use this

Checking calendars, tickets, or inventory mid-chat
Safe math via a calculator tool
“Book / fetch / search” actions behind your auth

Watch out

Validate arguments. Don’t let freeform strings delete files or spend money without guards. If the tool fails, say it failed — don’t let the model invent a success.

Try next

Try this next

Send a user message that should not call a tool. Confirm your dispatcher stays quiet.