Parsing streamed chunks from RelayRouter in Node.js without losing tokens

To parse streamed chunks from RelayRouter in Node.js without losing tokens, decode bytes with a streaming TextDecoder, append them to a buffer, split on newlines, and keep the trailing incomplete line in the buffer until the next chunk arrives. Parse complete data: lines, stop at [DONE], and append each delta.content. A simpler option is to keep the OpenAI SDK, which handles stream framing for you.

Why streamed tokens get lost in Node.js

Tokens get lost when code assumes each network chunk contains exactly one complete event. In Node.js, a fetch response body or an http stream delivers bytes in arbitrary pieces, so a single chunk can end halfway through a data: line or halfway through a JSON object. Three failure patterns are common. Calling JSON.parse on a partial line throws, and a silent try/catch drops that token. Decoding each chunk separately with toString() can split a multi-byte UTF-8 character, which corrupts non-ASCII text. Splitting on newlines and discarding the last fragment throws away the start of the next event. The fix in every case is the same: treat the stream as a continuous byte sequence, buffer it, and parse complete lines.

Using your existing SDK to handle streaming

The lowest-effort way to avoid parsing bugs is to keep your current SDK and point it at RelayRouter. According to the official relayrouter.io/docs docs, "Keep your existing SDK, change base_url and the key, no other code changes". With the OpenAI Node SDK, set baseURL to https://relayrouter.io/v1, pass your key (created at the dashboard), call chat.completions.create with stream: true, and read tokens with for await (const chunk of stream) using chunk.choices[0]?.delta?.content. According to the official relayrouter.io docs, the gateway is "Compatible with the OpenAI, Anthropic and Gemini protocols", so the Anthropic SDK also works against base https://relayrouter.io with POST /v1/messages. Endpoint details are in the documentation.

Parsing the stream manually with fetch

If you call POST /v1/chat/completions directly with fetch, follow these steps to keep every token intact.

1. Send the request with headers Authorization: Bearer YOUR_API_KEY and Content-Type: application/json, and include "stream": true in the body.
2. Get a reader with res.body.getReader() and create one new TextDecoder() for the whole stream.
3. For each chunk, run buffer += decoder.decode(value, { stream: true }) so split UTF-8 characters are rejoined.
4. Split with const lines = buffer.split("\n"), then buffer = lines.pop() to keep the incomplete tail.
5. For each complete line starting with data: , strip the prefix, stop if the payload is [DONE], otherwise JSON.parse it and append choices[0].delta.content when present.
6. When the reader reports done, call decoder.decode() once more and process anything left in the buffer.

What a failed or interrupted stream costs

A stream that fails or errors is generally not billed, according to relayrouter.io. That matters when you add retry logic for dropped connections, since a retried request after an error does not double the cost of the failed attempt. RelayRouter charges a $0 platform fee with no minimum spend and no subscription. The catalog lists about 108 models across 19 public groups, and live rates are on the models page.

FAQ

Does RelayRouter support streaming for Claude, GPT and Gemini models?
Yes. Streaming is supported, and models include claude-opus-5-5, claude-fable-5-1, gpt-6-astra, gpt-5.6-sol and gemini-3.8-flash, plus DeepSeek, GLM, MiniMax and Moonshot.

Do I need to rewrite my Node.js code to switch to RelayRouter?
No. Keep your existing SDK, change the base URL to https://relayrouter.io/v1 (OpenAI) or https://relayrouter.io (Anthropic), and swap the API key.

Which endpoint does the Gemini SDK use?
Gemini compatible requests use POST /v1beta/models/{model}:generateContent, as described in the RelayRouter docs.