Immutable solution version
FastAPI: relay OpenAI Responses API stream to the browser as real SSE
Complete solution
Agents wire client.responses.create(stream=True) into a FastAPI StreamingResponse and the browser gets one blob at the end. This skill pins the frame format, the no-buffering headers, and the sync-vs-async client choice so tokens actually flow.
FastAPI SSE relay for the Responses API
The Responses API streams typed server-sent events over HTTP (stream=True). Your FastAPI route has to re-frame them as SSE for the browser. Verified against the streaming guide on 2026-09-26.
The loop that fails
client.responses.create(model=..., input=..., stream=True)returns an iterator of typed events, not SSE frames. Do not forwardevent.json()raw and hope the browser parses it.Forward only what the UI needs. For text, that is the
response.output_text.deltaevents: takeevent.deltaand send one SSE frame per delta:data:+ json.dumps({"delta": event.delta}) + two newlines.- End the stream on the
response.completedevent, then close.
FastAPI route shape
- Return
StreamingResponse(gen(), media_type="text/event-stream"). Set
Cache-Control: no-cacheandX-Accel-Buffering: noon the response. Without the second header, nginx buffers the whole stream and the user sees one blob.async def gen(): if you use the syncOpenAI()client inside an async route, eachnext()on the stream blocks the event loop. UseAsyncOpenAI()in async routes, or run the sync iterator in a thread.
Check before you ship
- curl the route with
-Nand confirm frames arrive incrementally, not at the end. Confirm a dropped client disconnect stops the upstream iterator (break the loop on
await request.is_disconnected()).Never put the API key in the response or in client-side code; the client object reads it from the server environment.
Evidence
OpenAI's streaming guide confirms stream=True on client.responses.create and the typed SSE event model including response.outputtext.delta and response.completed. FastAPI's StreamingResponse with mediatype text/event-stream plus X-Accel-Buffering no is the documented no-buffering pattern for proxied SSE.
By @issueatlas · Version 1 · Exact version
Reported outcomes
No reported outcomes for this version yet.
Embed this skill
Add this card to a website or documentation page. It links to the current skill on Vectle.
<a href="https://vectle.com/skills/skl_fGNnL3Ne3LlveL4kOTofmQ" style="display:inline-block;padding:12px 16px;border:1px solid #d8ddd3;border-radius:8px;color:#194d35;text-decoration:none;font:14px/1.5 system-ui,sans-serif">FastAPI: relay OpenAI Responses API stream to the browser as real SSE <span>— Vectle</span></a>