Streaming

assistant-ui Streaming

Always consult assistant-ui.com/llms.txt for the latest API.

assistant-stream is the wire layer underneath assistant-ui's chat runtimes. It normalizes every backend into one stream of AssistantStreamChunk values, ships encoders and decoders for three wire formats, and adds a resumable-stream layer on top of any of them. If your backend already speaks the Vercel AI SDK, you rarely touch this package directly (streamText plus toUIMessageStream is enough); reach for it when you write a custom endpoint, need to decode a stream yourself, or want resumable streams.

References

• ./references/data-stream.md -- the Data Stream protocol, useDataStreamRuntime, and its wire format
• ./references/assistant-transport.md -- the Assistant Transport SSE format and the useAssistantTransportRuntime state-snapshot runtime
• ./references/encoders.md -- the encoder and decoder catalog, PlainTextEncoder, UIMessageStreamDecoder, accumulators, and debugging
• ./references/resumable.md -- assistant-stream/resumable: context, stores, and client wiring

When to use it

Streaming the model call through the Vercel AI SDK?
├─ Yes → streamText + toUIMessageStream/createUIMessageStreamResponse (or result.toUIMessageStreamResponse())
│        assistant-stream is optional: only needed to decode the response yourself or add resumable streams
└─ No → build the response with assistant-stream
    ├─ Emitting message parts (text, reasoning, tool calls) → Data Stream
    └─ Streaming a full agent state snapshot with custom commands → Assistant Transport

Installation

npm install assistant-stream

@assistant-ui/ai-sdk is the current AI SDK integration package (framework neutral); @assistant-ui/react-ai-sdk still re-exports the same API for older installs but new code should import from @assistant-ui/ai-sdk.

Build a custom streaming response

createAssistantStreamResponse runs a callback with an AssistantStreamController and returns a Response encoded as Data Stream (see data-stream.md for the alternative encoders).

import { createAssistantStreamResponse } from "assistant-stream";

export async function POST(req: Request) {
  return createAssistantStreamResponse(async (controller) => {
    controller.appendText("Hello ");
    controller.appendText("world!");

    controller.appendReasoning("Checking the forecast first.", {
      unstable_summary: "Looking up the weather",
    });

    controller.appendSource({
      type: "source",
      sourceType: "url",
      id: "s1",
      url: "https://example.com/forecast",
      title: "Forecast",
    });

    const tool = controller.addToolCallPart({ toolName: "get_weather" });
    tool.argsText.append('{"city":"NYC"}');
    tool.argsText.close();
    tool.setResponse({ result: { temperature: 22 } });

    controller.close();
  });
}

close() closes any part still open and ends the stream; an uncaught throw inside the callback is turned into an error chunk automatically.

AssistantStreamController

Every server-side stream, whichever encoder ends up wrapping it, is written through this controller (createAssistantStream, createAssistantStreamController, and createAssistantStreamResponse all hand you one).

addToolCallPart returns a ToolCallStreamController: { argsText: TextStreamController, setResponse(response), close() }. setResponse takes { result, artifact?, isError?, modelContent?, messages? } (the shape returned by a ToolResponse), closes the part automatically, and ignores a second call.

Stream events and part types

Every decoder, regardless of wire format, yields the same normalized AssistantStreamChunk union ({ path: number[] } & { type, ... }):

PartInit (the part field of part-start) is one of six part types, every variant carrying an optional parentId:

Common Gotchas

appendSource, appendFile, or appendData silently drops the part

• Pass the full part object including its type field ("source", "file", or "data"); the method name does not imply it for you.

A tool call never settles in the UI

• addToolCallPart needs a toolName; the id is generated for you unless you pass one. Close argsText (or call setResponse, which closes it for you) or the part never finishes. Register the rendering with a "use generative" toolkit, not the deprecated makeAssistantToolUI; see tools.

Two separate reasoning parts merge into one on the client

• On the Data Stream wire, a reasoning part-start frame is only sent when unstable_summary is set; a plain appendReasoning(text) call travels only as text deltas, and the decoder has nothing else to tell it a new part started. Opening two summary-less reasoning parts back to back (for example around a tool call) reconstructs as one continuous reasoning part on the client. Give each part a unstable_summary (even an empty-feeling one) or route the tool call through a separate message step to keep them distinct.

Stream not updating the UI

• Check the Content-Type against the encoder you actually used: DataStreamEncoder (the createAssistantStreamResponse default) sends text/plain; charset=utf-8 with x-vercel-ai-data-stream: v1, not text/event-stream. AssistantTransportEncoder and the AI SDK's UI message stream do send text/event-stream.

Decoder throws "Stream ended abruptly without receiving [DONE] marker"

• AssistantTransportDecoder and UIMessageStreamDecoder require the terminal [DONE] sentinel; a proxy, CDN, or middleware that buffers or truncates the body breaks this. DataStreamDecoder has no such marker.

createAssistantStreamResponse always encodes as Data Stream

• It hard-codes DataStreamEncoder. For a different wire format, encode manually: AssistantStream.toResponse(createAssistantStream(callback), new AssistantTransportEncoder()), or use createAssistantStreamController and encode the returned stream yourself.

Related Skills

• runtime -- useLocalRuntime, useExternalStoreRuntime, and the useAssistantTransportRuntime React hook and state hooks
• setup -- scaffolding an AI SDK route handler and useChatRuntime
• tools -- "use generative" toolkits and tool-call rendering
• cloud -- persisting streamed threads and messages with assistant-cloud