Skip to content
bits
Sign inStart free
Give your model tools · Lesson 10 of 10

Checkpoint: a weather bot

Tutor:

Time to put the course together. You're building a weather bot in TypeScript: one question in, one answer out, backed by real weather data. It uses everything you've learned. A tool with a description that says when not to use it. A strict schema. One round trip. Every call answered, in one message. Errors sent back with is_error.

The weather comes from Open-Meteo, which is free for non-commercial use and needs no key. You only need an Anthropic API key in ANTHROPIC_API_KEY.

import Anthropic from "@anthropic-ai/sdk";
const client = new Anthropic(); // reads ANTHROPIC_API_KEY
const tools: Anthropic.Tool[] = [
{
name: "get_weather",
description:
"Get the current weather for one city: temperature in °C, the sky and the wind. " +
"Use it whenever the user asks about weather, temperature or what to wear outdoors " +
"right now, once for each city. It has no forecasts or history, so don't use it " +
"for tomorrow or last week.",
strict: true,
input_schema: {
type: "object",
properties: {
city: {
type: "string",
description: "City name, with the country if it's ambiguous, e.g. Paris, France",
},
},
required: ["city"],
additionalProperties: false,
},
},
];
const SYSTEM =
"You are a friendly weather assistant. Answer in one or two sentences. " +
"Use get_weather for current conditions and never guess the weather.";
// WMO weather codes (Open-Meteo docs), the common ones.
const SKY: Record<number, string> = {
0: "clear", 1: "mainly clear", 2: "partly cloudy", 3: "overcast", 45: "fog",
51: "drizzle", 61: "light rain", 63: "rain", 65: "heavy rain",
71: "light snow", 73: "snow", 75: "heavy snow", 80: "showers", 95: "thunderstorm",
};
// The tool: Open-Meteo's free API, no key needed (non-commercial use).
async function getWeather(city: string) {
const geo = await fetch(
`https://geocoding-api.open-meteo.com/v1/search?count=1&name=${encodeURIComponent(city)}`,
);
const place = (await geo.json()).results?.[0];
if (!place) {
throw new Error(`No city called "${city}". Check the spelling, or try a nearby larger city.`);
}
const res = await fetch(
`https://api.open-meteo.com/v1/forecast?latitude=${place.latitude}` +
`&longitude=${place.longitude}&current=temperature_2m,weather_code,wind_speed_10m`,
);
if (!res.ok) throw new Error(`The weather service failed (HTTP ${res.status}). Try again later.`);
const { current } = await res.json();
return {
city: `${place.name}, ${place.country}`,
temp_c: current.temperature_2m,
sky: SKY[current.weather_code] ?? "mixed",
wind_kmh: current.wind_speed_10m,
};
}
function printText(reply: Anthropic.Message) {
for (const block of reply.content) if (block.type === "text") console.log(block.text);
}
const question = process.argv[2] ?? "Do I need a coat in Chicago?";
const messages: Anthropic.MessageParam[] = [{ role: "user", content: question }];
const ask = () =>
client.messages.create({ model: "claude-haiku-5-5", max_tokens: 4096, system: SYSTEM, tools, messages });
// Request one: Claude reads the question and the tool, and decides.
const first = await ask();
if (first.stop_reason !== "tool_use") {
printText(first); // no tool needed
process.exit(0);
}
// Run every call (there may be several), and answer every call.
const calls = first.content.filter((block) => block.type === "tool_use");
const results = await Promise.all(
calls.map(async (call): Promise<Anthropic.ToolResultBlockParam> => {
const { city } = call.input as { city: string };
console.error(`→ get_weather("${city}")`);
try {
const weather = await getWeather(city);
return { type: "tool_result", tool_use_id: call.id, content: JSON.stringify(weather) };
} catch (error) {
const message = error instanceof Error ? error.message : String(error);
return { type: "tool_result", tool_use_id: call.id, content: message, is_error: true };
}
}),
);
// Request two: Claude's reply as it came, then every result in one message.
messages.push({ role: "assistant", content: first.content }, { role: "user", content: results });
const second = await ask();
printText(second);
if (second.stop_reason === "tool_use") {
console.error("(Claude wants another call. Looping comes in Your first agent loop.)");
}
TypeScriptUTF-8● lines 71–100: the round trip
OUTPUT
npx tsx weather-bot.ts "Is it warmer in Paris or Tokyo right now?"
example output, written for this lesson (not a live run)
→ get_weather("Paris, France")
→ get_weather("Tokyo, Japan")
Tokyo is warmer right now: 18°C and partly cloudy, against 10°C and clear in Paris.
Tutor:

Read it top to bottom and you'll recognise every part. Two calls came back for Paris and Tokyo in one reply, ran side by side, and went back in one message.

The tool also returns the place it actually found, not just the temperature. That matters. Open-Meteo matches names loosely: ask it for “Pariss” and it finds Périsse, in Haiti. Because the result says "Périsse, Haiti", Claude can spot the mismatch and say so, instead of cheerfully reporting Caribbean weather for Paris.

quick check+20 neurons
Tutor:

Someone asks about the weather in “Xyzzyville”. Open-Meteo finds nothing. What does the bot do?

    Type your answer or tap an option · Enter checks

    quick check+20 neurons
    Tutor:

    Someone asks “Will it rain in Leeds tomorrow?”. What should a good run look like?

      Type your answer or tap an option · Enter checks

      Tutor:

      Last step. The description is the part that decides everything, so test yours. Write a get_weather description that covers what it does, when to use it and what it can't do, then ask Claude to say tool or no tool for the three messages in the draft. The sandbox has no tools, so Claude is telling you its decision rather than making the call. It's the same judgement, from the same description.

      Your turn

      Run your prompt and the replies land here.

        • Names the tool, not yet
        • Says what it does, not yet
        • Says when to use it, not yet
        • Says what it can't do, not yet
        • Keeps the three messages, not yet
        • Ran at 1.0, not yet
        run at
        ~0 tokens

        Turn on JavaScript to run prompts.

        Tutor:

        That's a working tool-using bot. It answers one question with at most one round of tool calls. If Claude wants a second round (look up the city, then the weather, then a packing list), the bot stops and says so.

        Handling that is a later course, Your first agent loop. You'll wrap this round trip in a loop that keeps going until Claude has what it needs, and learn how to stop it when it doesn't.

        Down the rabbit hole