LLM SSE Stream Decoder

Paste raw SSE bytes from an OpenAI or Anthropic streaming response and instantly decode the assembled text, finish/stop reason, and a per-event breakdown. Supports OpenAI chat completions, OpenAI responses API, and Anthropic Messages streaming — select the format before running.

How to use the LLM SSE Stream Decoder

Paste the raw bytes of an SSE response — exactly as you would see in a terminal using curl -N — into the top textarea. Each line should start with data: followed by a JSON object, or be data: [DONE] for OpenAI streams. Select the API format that matches your stream source, then click Decode.

The tool outputs:

  • The assembled message text — all deltas concatenated in order.
  • The finish/stop reason — e.g. stop, length, tool_calls for OpenAI; end_turn, max_tokens for Anthropic.
  • A per-event table showing the index, event type, and delta content for each parsed event.

Use the Example button to load a realistic sample stream for the selected format and see the decoder in action before pasting your own data.

Server-Sent Events and LLM streaming

Server-Sent Events (SSE) is the W3C standard used by OpenAI and Anthropic to stream partial completions to clients in real time. The wire format is simple: each event is a data: prefixed line of text, events are separated by blank lines, and a stream ends with either data: [DONE] (OpenAI) or a terminal event like message_stop (Anthropic). The browser EventSource API and HTTP client libraries handle the framing; debugging requires reassembling the raw bytes manually.

Each API has its own JSON schema for delta events. OpenAI chat completions place the token delta at choices[0].delta.content and signal the end with a separate data: [DONE] sentinel. The newer Responses API emits named event types (response.output_item.added, response.text.delta). Anthropic emits events with a top-level type field; text deltas arrive as content_block_delta events with delta.type === "text_delta" and delta.text, while message_stop terminates the stream and message_delta carries the stop reason.

Common debugging scenarios: a stream that terminates early (max_tokens hit), a stream where tool_calls appear instead of text, or verifying that all delta chunks were received in order after a network interruption. This decoder surfaces all three.

Common use cases

  • Debugging truncated responses — paste a stream to see exactly where it stopped and what the finish reason was.
  • Verifying streaming integration — confirm your client code reassembles deltas in the correct order.
  • Comparing API formats — see side-by-side how OpenAI and Anthropic structure their stream events.
  • Inspecting tool call streams — decode streams where the model is generating function call arguments incrementally.
  • Learning the SSE protocol — use the event table to understand exactly which event types fire in what order.

Frequently asked questions

Why does my stream end with data: [DONE] and no text?
This usually means the model returned a tool_calls delta instead of a content delta. The assistant is calling a function rather than generating text. Check the event table for a delta with role assistant and tool_calls content.

What is the difference between OpenAI Chat and Responses API formats?
Chat Completions wraps deltas in a choices[0].delta object and uses data: [DONE] to end. The Responses API (for the newer /v1/responses endpoint) uses named event types like response.text.delta and response.completed instead.

How do Anthropic streams differ from OpenAI?
Anthropic uses a top-level type field on every event. Text arrives in content_block_delta events with delta.text. The stream ends with a message_stop event; the stop_reason appears in a preceding message_delta event. There is no [DONE] sentinel.

Can I debug tool call argument streams?
Yes — for OpenAI, tool call argument chunks appear in choices[0].delta.tool_calls[0].function.arguments. The decoder will show these in the event table; the assembled text output will be empty since there are no content deltas.