SSE & chunked, passed through.
Server-sent events or chunked transfer, forwarded to your caller as your target produces them — no waiting for the whole response.
Streaming
Stream it straight back.
Point a streaming target at EchoRelay and we forward its SSE or chunked response to your caller, chunk by chunk — built for LLM token streaming.
Built for AI
Tokens, chunk by chunk.
Made for model output.
Stream an LLM’s tokens or a long-running inference job straight to the caller — incremental, not buffered to the end.
Configured by your agent, over MCP.
Set up the streaming endpoint with a sentence to your agent — it speaks EchoRelay’s MCP surface and ships the config for you.
One stream per endpoint.
A streaming endpoint carries a single stream target — clear, predictable behaviour.
How it works
Open, forward, stream.
-
Your caller opens a request
A client hits your streaming endpoint and holds the connection open, waiting to receive.
-
We authenticate and open the stream
We verify your key and apply rate limits at the edge, then open the connection to your streaming target.
-
Chunks stream straight back
As your target emits SSE or chunked output, we forward each chunk to your caller as it arrives — incremental, not buffered to the end.
Streaming runs through the same durable relay as everything else — your caller gets the stream while we keep auth, validation, and rate limits at the edge. A stream is billed by the size of the response it streams back — heartbeats are free, and a cut-short stream only bills for what was delivered.
Questions
Streaming, answered.
- Does EchoRelay buffer the whole response before sending it?
- No. Chunks are forwarded as your target produces them, so your caller sees output as it is generated rather than waiting for the full response.
- What happens if a stream is cut short?
- Your caller keeps everything delivered up to the cut, and billing only counts what was actually streamed back. Heartbeats are not charged.
- Can a streaming endpoint fan out to several targets?
- No — a streaming endpoint carries a single stream target, for clear and predictable behaviour. Delivering one request to many targets is a separate shape.
- Is streaming only for LLMs?
- No. Any SSE or chunked HTTP response works. LLM token streaming is the common case, but long-running inference and progress streams behave the same way.
Not the right shape? To deliver one request to many targets at once, use Fan-out. To take inbound traffic while keeping your origin off the public internet, use Origin shield.
Put us between your model and your users.
No credit card required.