Unstyled React primitives for generative/streaming UI patterns — compatible with Vercel AI SDK, Anthropic, OpenAI, and more.
npm install @raed667/streaming-ui-primitives
# or
pnpm add @raed667/streaming-ui-primitives
# or
yarn add @raed667/streaming-ui-primitivesPeer dependencies: React 18+
import { useTokenStream, StreamingText, StreamGuard } from '@raed667/streaming-ui-primitives'
import { fromFetchSSE } from '@raed667/streaming-ui-primitives/adapters'
function ChatMessage() {
const [stream, setStream] = React.useState<AsyncIterable<string> | null>(null)
const { text, status, error } = useTokenStream(stream)
async function ask(prompt: string) {
const res = await fetch('/api/chat', {
method: 'POST',
body: JSON.stringify({ prompt }),
})
setStream(fromFetchSSE(res))
}
return (
<div>
<button onClick={() => ask('Hello!')}>Send</button>
<StreamGuard
status={status}
idle={<p>Ask me anything...</p>}
streaming={<StreamingText content={text} isStreaming cursor />}
complete={<StreamingText content={text} isStreaming={false} />}
error={(err) => <p>Error: {err?.message}</p>}
/>
</div>
)
}Streaming markdown — PartialRender with a markdown renderer, token by token:
Typing indicator — three built-in variants, unstyled (inherits color):
Streaming text with cursor — StreamingText with blinking cursor while active:
Consumes any token source and accumulates the text, tracking the full streaming lifecycle.
import { useTokenStream } from '@raed667/streaming-ui-primitives'
// source can be AsyncIterable<string>, ReadableStream<Uint8Array>, or null/undefined
const { text, isStreaming, status, error, abort, reset } = useTokenStream(source)Pass null or undefined to keep the hook idle. When source changes, the previous stream is automatically aborted.
Bridges Vercel AI SDK's UIMessage.parts array to simple derived values. No runtime dependency on the ai package.
import { useMessageStream } from '@raed667/streaming-ui-primitives'
const { messages, status } = useChat({ api: '/api/chat' })
const lastMsg = messages[messages.length - 1]
const { text, reasoning, hasActiveToolCall, hasReasoning, sourceUrls } = useMessageStream(
lastMsg?.parts ?? [],
)Debounces isStreaming to prevent flicker when a stream pauses briefly between tokens.
import { useDebouncedStreaming } from '@raed667/streaming-ui-primitives'
const { isStreaming } = useTokenStream(source)
const stableStreaming = useDebouncedStreaming(isStreaming, 150) // default: 150msMaps a Vercel AI SDK useChat status string directly to StreamStatus.
import { useAISDKStatus } from '@raed667/streaming-ui-primitives'
const { status } = useChat({ api: '/api/chat' })
const streamStatus = useAISDKStatus(status) // 'idle' | 'streaming' | 'complete' | 'error'Renders text that grows token-by-token. Completely unstyled — no fonts, colors, or layout applied.
import { StreamingText } from '@raed667/streaming-ui-primitives'
// Blinking cursor while streaming, hidden when complete
<StreamingText content={text} isStreaming={isStreaming} cursor />
// Custom cursor node
<StreamingText content={text} isStreaming={isStreaming} cursor={<span>▋</span>} />
// Render as a <p> instead of the default <span>
<StreamingText content={text} as="p" className="prose" />Animated "AI is thinking" indicator. Inherits color from the parent — fits any theme automatically.
import { TypingIndicator } from '@raed667/streaming-ui-primitives'
// Three bouncing dots (default)
<TypingIndicator visible={isStreaming} />
// Single pulsing circle
<TypingIndicator visible={isStreaming} variant="pulse" />
// Three animated vertical bars
<TypingIndicator visible={isStreaming} variant="bar" />
// Custom accessible label
<TypingIndicator visible={isStreaming} aria-label="Generating response..." />Gracefully renders partial/incomplete content during streaming. Uses an error boundary to catch parse errors mid-stream and fall back to plain text. Renderer-agnostic — pass any render function.
import { PartialRender } from '@raed667/streaming-ui-primitives'
import { marked } from 'marked'
;<PartialRender
content={text}
isComplete={status === 'complete'}
renderer={(content, isComplete) => <div dangerouslySetInnerHTML={{ __html: marked(content) }} />}
fallback={<span>Thinking...</span>}
errorFallback={(err, content) => <pre>{content}</pre>}
/>Status-driven render slots — a type-safe switch/case over stream lifecycle states.
import { StreamGuard } from '@raed667/streaming-ui-primitives'
;<StreamGuard
status={status}
idle={<p>Ask me anything...</p>}
streaming={<TypingIndicator visible />}
complete={<StreamingText content={text} />}
error={(err) => <p>Something went wrong: {err?.message}</p>}
errorValue={error}
/>Adapters are in a separate entry point to keep the core bundle lean:
import {
fromFetchSSE,
fromOpenAIChatStream,
fromAnthropicStream,
partsToText,
hasActiveToolCall,
} from '@raed667/streaming-ui-primitives/adapters'Converts a raw fetch Response body (plain text or Server-Sent Events) into an AsyncIterable<string>.
// Plain text stream (auto-detected from Content-Type)
const res = await fetch('/api/stream')
const stream = fromFetchSSE(res)
// SSE stream with `data:` prefix (Vercel AI SDK text stream, OpenAI-compatible endpoints)
const res = await fetch('/api/chat', {
method: 'POST',
body: JSON.stringify(payload),
})
const stream = fromFetchSSE(res, { mode: 'sse-text' })
// SSE stream with JSON payloads — custom dot-path extraction
const stream = fromFetchSSE(res, {
mode: 'sse-json',
jsonPath: 'choices.0.delta.content',
})
// Vercel AI SDK data stream protocol (bare `0:"token"` lines, no `data:` prefix)
const stream = fromFetchSSE(res, { mode: 'vercel-ai' })
// Use with useTokenStream inside a component
const [stream, setStream] = useState(null)
const { text } = useTokenStream(stream)
async function send(prompt) {
const res = await fetch('/api/chat', {
method: 'POST',
body: JSON.stringify({ prompt }),
})
setStream(fromFetchSSE(res))
}Converts an Anthropic SDK client.messages.stream(...) result to AsyncIterable<string>. Zero runtime dependency — accepts the shape structurally.
import Anthropic from '@anthropic-ai/sdk'
import { fromAnthropicStream } from '@raed667/streaming-ui-primitives/adapters'
const client = new Anthropic()
const stream = client.messages.stream({
model: 'claude-opus-4-5',
max_tokens: 1024,
messages: [{ role: 'user', content: prompt }],
})
// Option A — adapter (handles raw Stream<RawMessageStreamEvent>)
const { text } = useTokenStream(fromAnthropicStream(stream))
// Option B — direct (stream.textStream is already AsyncIterable<string>)
const { text } = useTokenStream(stream.textStream)Converts an OpenAI SDK chat.completions.stream(...) result to AsyncIterable<string>. Zero runtime dependency on the OpenAI SDK — accepts the shape structurally.
import OpenAI from 'openai'
import { fromOpenAIChatStream } from '@raed667/streaming-ui-primitives/adapters'
const openai = new OpenAI()
const stream = openai.chat.completions.stream({
model: 'gpt-4o',
messages: [{ role: 'user', content: prompt }],
stream: true,
})
const tokenStream = fromOpenAIChatStream(stream)
// In a component:
const { text, isStreaming } = useTokenStream(tokenStream)Same as above but for legacy text completions (openai.completions.create(..., { stream: true })).
const stream = openai.completions.create({
model: 'gpt-3.5-turbo-instruct',
prompt,
stream: true,
})
const tokenStream = fromOpenAICompletionStream(stream)Extracts plain text from a Vercel AI SDK UIMessage.parts array. Concatenates all type: 'text' parts in order.
import { partsToText } from '@raed667/streaming-ui-primitives/adapters'
const { messages } = useChat({ api: '/api/chat' })
const lastMessage = messages[messages.length - 1]
const text = partsToText(lastMessage.parts)
// Include reasoning parts too
const textWithReasoning = partsToText(lastMessage.parts, {
includeReasoning: true,
})Returns true if any tool-invocation part is in a non-result state.
import { hasActiveToolCall } from '@raed667/streaming-ui-primitives/adapters'
const isToolRunning = hasActiveToolCall(message.parts)
<TypingIndicator visible={isToolRunning} aria-label="Running tool..." />Utility function (also exported from the main entry) that maps a Vercel AI SDK useChat status to StreamStatus.
import { fromUseChatStatus } from '@raed667/streaming-ui-primitives'
const { status } = useChat({ api: '/api/chat' })
const streamStatus = fromUseChatStatus(status)
// 'submitted' | 'streaming' → 'streaming'
// 'ready' → 'complete'
// 'error' → 'error'| SDK | How to connect | Notes |
|---|---|---|
| Vercel AI SDK v4+ | fromUseChatStatus(status), useMessageStream(message.parts), or fromFetchSSE(res, { mode: 'sse-text' }) |
useChat status maps via fromUseChatStatus; parts are handled by useMessageStream / partsToText |
| Anthropic SDK | Pass stream.textStream directly to useTokenStream(), or use fromAnthropicStream(stream) |
stream.textStream is AsyncIterable<string> — no adapter needed; fromAnthropicStream handles raw Stream<RawMessageStreamEvent> |
| OpenAI SDK v4+ | fromOpenAIChatStream(stream) |
Wraps chat.completions.stream(); use fromOpenAICompletionStream for legacy completions |
| LangChain.js | Pass chain.stream() directly to useTokenStream() |
Any AsyncIterable<string> works without an adapter |
| LlamaIndex.TS | Pass engine.chat({ stream: true }) result directly to useTokenStream() |
Any AsyncIterable<string> works without an adapter |
| Raw fetch SSE | fromFetchSSE(response) |
Auto-detects plain text vs SSE from Content-Type; use mode option to force |
Five standalone runnable apps in the examples/ folder, each a Vite + React SPA with an Express mock server:
| Folder | SDK | Patterns |
|---|---|---|
examples/vercel-ai/ |
Vercel AI SDK | useMessageStream, fromUseChatStatus, StreamGuard, TypingIndicator, useDebouncedStreaming, multi-turn history |
examples/anthropic/ |
Anthropic SDK | fromAnthropicStream, StreamingText with cursor, abort(), StreamGuard error state |
examples/openai/ |
OpenAI SDK | fromOpenAIChatStream, StreamGuard, abort mid-stream, multi-turn chat |
examples/raw-fetch/ |
None (raw fetch) | fromFetchSSE all 4 modes: auto, sse-text, vercel-ai, sse-json, reset() |
examples/kitchen-sink/ |
None (mock) | Markdown rendering, custom cursor, TypingIndicator variants, abort vs reset, debounced streaming, reasoning, source URLs, StreamGuard all 4 states |
To run any example:
cd examples/<name> # e.g. examples/vercel-ai
pnpm install
pnpm dev # starts Vite on :5173 + Express mock server on :3001Then open http://localhost:5173. No API key needed — each example ships with a mock server that streams fake tokens. See examples/README.md for details on using real AI SDKs.
# Run unit tests
pnpm test
# Run end-to-end tests (builds Storybook first)
pnpm test:e2e:ci