Meet Claude — the API model

You talk to Claude by sending an HTTP POST to the /v1/messages endpoint. The body includes a model name, a messages array (the conversation so far), and parameters like max_tokens. Claude responds with an assistant message. That's the whole API surface. Everything sophisticated is built on top of these two primitives: send messages, receive messages.

A conversation is an ordered array of message objects. Each has a role and content. The array is stateless — YOU maintain the conversation history and re-send it every call. There is no server-side session. This means to have a multi-turn conversation, you append each new turn (user OR assistant) to the array and send the whole thing again next time.

System (a top-level parameter, not a message role): persistent instructions and role for Claude. User: what a human said (or a tool result). Assistant: what Claude said (or wants to do next). You'll only ever write user messages; assistant messages come back from the API and you store them in the array.

Here's the simplest possible call to Claude. In practice you'd use the anthropic Python SDK, but this is the raw shape.

import anthropic

client = anthropic.Anthropic()  # reads API key from env

response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system="You are a helpful assistant.",
    messages=[
        {"role": "user", "content": "What's the capital of France?"}
    ]
)

print(response.content[0].text)
# → "The capital of France is Paris."

Claude's response includes: content (an array of content blocks — usually text, sometimes tool_use), stop_reason (why the model stopped — end_turn, tool_use, max_tokens, stop_sequence), usage (input and output token counts, for cost tracking), and model / id metadata. The content is an array because a single response can mix text and tool_use blocks.

model: which Claude version. max_tokens: cap on output length (required). temperature: 0 to 1, controls randomness. system: the persistent system prompt. tools: tool definitions (for function calling). tool_choice: force a specific tool. stop_sequences: strings that halt generation early.