Streaming responses

Enable streaming

Set stream to true. Ask for the final usage frame when your client supports it.

curl -N "$MAXLABS_BASE_URL/chat/completions" \
  -H "Authorization: Bearer $MAXLABS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "worker",
    "messages": [{"role": "user", "content": "Write a release checklist."}],
    "stream": true,
    "stream_options": {"include_usage": true}
  }'

The response uses Server-Sent Events. Each event begins with data: and contains a JSON object. The stream ends with this marker:

data: [DONE]

Join choices[0].delta.content values in order. Tool call arguments can also arrive across several frames, so concatenate them before parsing the final JSON arguments.

Handle errors inside the stream

A streaming connection can start with HTTP 200 and later emit an error object. Check every JSON frame for error; do not treat [DONE] alone as proof of success. If the connection closes early, record the request ID and apply the retry guidance.

Do not assume that cancelling the client stops account usage immediately. A request that has already started may continue long enough to settle its actual usage.