Streaming replies

Relay accepts the output your agent already produces. Send one request with stream=true and pipe a Vercel AI SDK UIMessageStream v1 into its body. Relay consumes the whole stream and commits one canonical message when it finishes. There is no Relay draft to open, append, edit, or clean up.

Pipe an AI SDK response

Use the conversation_id from message.received and derive the idempotency key from that event’s event_id.

const result = streamText({ model, messages });
const uiResponse = result.toUIMessageStreamResponse();

const relayResponse = await fetch(
  `${process.env.RELAY_API_URL}/v1/messages?stream=true&conversation_id=${event.data.message.conversation_id}`,
  {
    method: "POST",
    headers: {
      Authorization: `Bearer ${process.env.RELAY_AGENT_TOKEN}`,
      "Idempotency-Key": `reply-${event.event_id}`,
      "Content-Type": "text/event-stream",
      "x-vercel-ai-ui-message-stream": "v1",
    },
    body: uiResponse.body,
    duplex: "half",
  },
);

if (!relayResponse.ok) throw new Error(await relayResponse.text());

This is a direct HTTPS integration, not a Relay SDK. Your backend still owns the model, tool loop, prompts, memory, and hosting.

Wire request

The same contract can be exercised without a framework:

curl -sS -X POST \
  "$RELAY_API_URL/v1/messages?stream=true&conversation_id=cnv_01JZC7K4RQ" \
  -H "Authorization: Bearer $RELAY_AGENT_TOKEN" \
  -H "Idempotency-Key: reply-evt_01JZE9M2XW" \
  -H "Content-Type: text/event-stream" \
  -H "x-vercel-ai-ui-message-stream: v1" \
  --data-binary @reply.sse

reply.sse uses the AI SDK’s existing framing:

data: {"type":"start","messageId":"ui-message-42"}

data: {"type":"text-start","id":"answer"}

data: {"type":"text-delta","id":"answer","delta":"Tomorrow at "}

data: {"type":"text-delta","id":"answer","delta":"2:00 PM works."}

data: {"type":"text-end","id":"answer"}

data: {"type":"finish","finishReason":"stop"}

data: [DONE]

Relay returns the normal 202 message response after the canonical commit.

What Relay preserves

AI SDK part	Relay presentation
Text parts	Canonical `text` parts
Tool input and output	A transcript `tool_call` data part; Relay never executes it
URL sources	`link_preview` parts
Files and document sources	Transcript artifact parts when they have usable metadata
Reasoning and transient custom data	Remain private to the agent backend

The external agent owns the full loop. Relay is the messaging channel and presentation surface.

Completion and recovery

finish is semantic completion. [DONE] only closes the transport; it does not commit a message by itself.
abort, error, malformed ordering, or an early disconnect writes no transcript row.
A successful stream produces one durable message.created event and one notification.
Retry the whole request with the same Idempotency-Key. The same completed stream returns the original message; different content returns 409 idempotency_conflict.
A disconnected app recovers the final message through normal cursor sync and history.

Use a separate typing indicator before the output stream begins when the agent has a long planning or tool phase.

Overview

Get started

Messages

Message components

Conversations

Webhooks & events

Reference

Streaming replies

Pipe an AI SDK response

Wire request

What Relay preserves

Completion and recovery

Next steps

​Pipe an AI SDK response

​Wire request

​What Relay preserves

​Completion and recovery

​Next steps

Pipe an AI SDK response

Wire request

What Relay preserves

Completion and recovery

Next steps