TypeScript clients

Core client

@streamweld/client has no runtime dependency. It supports browsers, Node.js, Deno, Bun, and edge runtimes with standard Fetch APIs.

npm install @streamweld/client

import { createDurableStream } from '@streamweld/client';

const stream = createDurableStream({
  url: 'https://inference.example.com/v1/chat/completions',
  body: {
    model: 'llama-8b',
    messages: [{ role: 'user', content: 'Count to five.' }],
    stream: true,
  },
  resume: { maxAttempts: 5 },
  onMigration(event) {
    console.debug('producer moved', event.fromBackend, event.toBackend);
  },
});

for await (const delta of stream.text) render(delta);
const outcome = await stream.result;

stream.text and stream.events are independent iterables over one multicast connection. Transport failure reconnects from the exact last processed event ID. Before the server reveals a stream ID, retries reuse one stable idempotency key so a lost initial response cannot create a second generation.

Call await stream.stop() only when the user intends to cancel generation. An AbortSignal detaches this reader and leaves the producer resumable. The stop request times out after 30 seconds, including its response body. A StreamTransportError does not establish whether the remote stop completed; retrying stop() is safe.

Persist across page reloads

import {
  createDurableStream,
  createLocalStoragePersistence,
} from '@streamweld/client';

const stream = createDurableStream({
  url,
  body,
  persist: createLocalStoragePersistence('current-generation'),
});

The helper preserves 64-bit cursor values as decimal strings. Applications with their own store can pass resumeFrom: { id, lastEventId }.

Vercel AI SDK v5

npm install @streamweld/ai-sdk ai@^5

The adapter installation notes cover strict TypeScript declaration dependencies and the tested npm dependency override for AI SDK v5.

For React's useChat, also install the v2 integration:

npm install @ai-sdk/react@^2

Change the useChat transport:

import { useChat } from '@ai-sdk/react';
import { StreamweldChatTransport } from '@streamweld/ai-sdk';

const chat = useChat({
  transport: new StreamweldChatTransport({
    api: 'https://inference.example.com/v1/chat/completions',
    model: 'llama-8b',
  }),
});

The adapter accepts text message parts and rejects unsupported tool, reasoning, and file parts before opening a request. Its abort signal detaches locally; use the adapter's explicit stop(chatId) operation for cancellation.

With StreamweldChatPersistence configured, stop(chatId) also works after a page reload. It stops the saved generation without replaying its response and keeps the checkpoint if the request fails, so you can retry.