SSE / Stream Response Replayer
Parse and replay raw LLM SSE logs from OpenAI, Anthropic, Ollama, and Vercel AI SDK
Hello world
No tool calls detected.
| # | Δt | Event | Description | Content so far |
|---|---|---|---|---|
| 0 | — | chunk | content.delta "Hel" | 3 chars |
| 1 | — | chunk | content.delta "lo" | 5 chars |
| 2 | — | chunk | content.delta " world" finish_reason=stop | 11 chars |
Frequently Asked Questions
What is the SSE Stream Replayer?
LLM APIs stream responses as Server-Sent Events (SSE): a sequence of `data: {...}` JSON chunks terminated by `data: [DONE]`. The replayer parses those raw logs (captured from curl, a proxy, or your backend) and reconstructs the full reply — content text, tool_calls, usage, finish reason — so you can debug what actually happened over the wire.
Which providers' formats does it understand?
OpenAI Chat Completions, OpenAI Responses API, Anthropic Messages (with `content_block_delta`, `content_block_start`, `message_delta` events), Ollama streaming, and the Vercel AI SDK's custom data stream protocol. Each has different chunk shapes, so the parser auto-detects and switches formats.
When do I actually need this?
- Your stream cut off mid-tool-call and you need to see which event was last - Token deltas look corrupted and you suspect a buffering / chunking bug in your client - You want to replay a captured stream to show a teammate or add it to a test fixture - A model refused or loops and you want the timeline of events
My log won't parse. What should I check?
Common causes: the log is missing the `data: ` prefix (some tools strip it), lines are stripped of newlines, or the capture includes HTTP headers / binary framing. Paste the raw SSE body between `HTTP/1.1 200 OK` and the closing `[DONE]` — the replayer tolerates missing `[DONE]` but requires proper `data:` lines.
Is my log data uploaded?
No. Parsing, timeline reconstruction, and typewriter replay run entirely in the browser. Since SSE logs often contain full prompts and responses with customer data, this matters — feel free to paste production captures.
How is this better than watching the Network tab?
Chrome's Network tab shows raw chunks but does not reconstruct tool_call deltas across events, normalize Anthropic vs OpenAI formats, or let you replay on a timeline. The replayer is purpose-built for LLM streaming shapes, including partial JSON assembly for tool arguments.