Streaming with Claude

Stream tokens as they arrive — critical for user-facing applications. Without streaming, users stare at a blank screen for 3–15 seconds. With streaming, they see the first token in <200ms.

Why Streaming Matters

Metric	No Streaming	With Streaming
Time to first byte (TTFB)	3-15s (full generation)	100-300ms
Perceived latency	Same as actual latency	Near-instant
User abandonment	40%+ after 5s	<5%
Memory on client	Buffer entire response	Process token by token

How SSE (Server-Sent Events) Works

# SSE Protocol — one-way server → client over HTTP
# 1. Client opens persistent HTTP connection
# 2. Server pushes events as they happen
# 3. Client processes each event immediately

# Claude API sends these SSE event types:
# message_start        → conversation metadata
# content_block_start  → new text or tool_use block
# content_block_delta  → incremental text token
# content_block_stop   → block complete
# message_delta        → stop_reason, usage
# message_stop         → done

Basic Text Streaming

with client.messages.stream(
    model="claude-sonnet-4-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Explain agents"}]
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

Streaming with Tool Use (Agent Loop)