Learning the Claude API: from API key to streaming

Claude is Anthropic’s AI model — the one I use in nearly every client AI feature: CS chatbots, RAG, and agents. This tutorial goes from getting an API key to streaming, using the official TypeScript/JavaScript SDK.

1. Setup

1. Create an account at platform.claude.com, then create an API key in the Console.
2. Store it as an environment variable — never put the key in code:

export ANTHROPIC_API_KEY="sk-ant-..."   # Windows (PowerShell): $env:ANTHROPIC_API_KEY="sk-ant-..."

3. Install the official SDK:

npm install @anthropic-ai/sdk

2. Your first call

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic(); // reads ANTHROPIC_API_KEY automatically

const response = await client.messages.create({
  model: "claude-opus-4-8",
  max_tokens: 16000,
  messages: [{ role: "user", content: "Explain RAG in 3 sentences, for an online store owner." }],
});

// content is an array of blocks — filter for text blocks before reading .text
for (const block of response.content) {
  if (block.type === "text") console.log(block.text);
}

Two important details: max_tokens is an upper bound on output (not a target), and response.content is an array of typed blocks — always check block.type first.

3. System prompts: personality & rules

const response = await client.messages.create({
  model: "claude-opus-4-8",
  max_tokens: 16000,
  system:
    "You are customer support for 'Gayo Store', a coffee shop. Answer warmly and concisely. " +
    "If asked about off-topic things, politely redirect. Never invent stock levels or prices.",
  messages: [{ role: "user", content: "Which arabica is your best seller?" }],
});

The system prompt is where business rules live — this is 80% of the work of building a good chatbot.

4. Multi-turn conversations

The API is stateless: send the full history every time.

const messages: Anthropic.MessageParam[] = [
  { role: "user", content: "My name is Budi, I sell coffee." },
  { role: "assistant", content: "Hi Budi! Nice to meet you." },
  { role: "user", content: "What's my name and what do I sell?" },
];

const response = await client.messages.create({
  model: "claude-opus-4-8", max_tokens: 16000, messages,
});

5. Streaming — mandatory for chat UIs

Without streaming, users stare at a blank screen until the answer finishes. With it, text flows word by word:

const stream = client.messages.stream({
  model: "claude-opus-4-8",
  max_tokens: 16000,
  messages: [{ role: "user", content: "Write a product description for 250g Gayo coffee." }],
});

stream.on("text", (delta) => process.stdout.write(delta));
const final = await stream.finalMessage();   // complete message + token usage
console.log("\nOutput tokens:", final.usage.output_tokens);

6. Choosing a model

Model	ID	When to use
Claude Opus 4.8	claude-opus-4-8	My default — strongest reasoning for agents & complex tasks
Claude Sonnet 5	claude-sonnet-5	The speed/cost balance for coding & chatbots
Claude Haiku 4.5	claude-haiku-4-5	High-volume light tasks: classification, routing

Per-token pricing differs by model — check the official pricing page; real costs only stay under control if you watch response.usage from day one.

Tips from experience

• Handle errors with the SDK’s typed exception classes (Anthropic.RateLimitError, etc.) — never string-match error messages.
• For store chatbots, answers must be grounded in data (RAG) — any model will hallucinate if asked to answer stock questions from memory. See the NexaMart case-study docs on this site.
• In production, keep the API key in a secret manager (on Cloudflare Workers: wrangler secret put), not in a committed .env file.