llm-stream-assemble: A TypeScript Solution for Unifying LLM Streaming Responses Across Multiple Platforms

llm-stream-assemble is a TypeScript module designed to address the differences in streaming API implementations among various LLM providers (such as OpenAI and Anthropic), unifying them into a standard event format to simplify multi-model integration development. The project is maintained by 01laky and was released on GitHub (link: https://github.com/01laky/llm-stream-assemble) on May 25, 2026.

In LLM application development, streaming responses have become a standard practice to enhance user experience, but there are significant differences in streaming API implementations among different model providers: OpenAI uses the SSE format with a delta field, Anthropic's Claude uses SSE but has a different event structure, and open-source model interfaces have subtle deviations. This fragmentation forces developers to write specialized parsing logic for each model. The llm-stream-assemble project was created to address this, providing a TypeScript layer module to uniformly convert these into a standardized event format.

llm-stream-assemble supports unifying the streaming responses of OpenAI, Anthropic, and compatible LLMs into standardized events, including: text content events (text), tool call events (tool calls), reasoning process events (reasoning). It also handles JSON structured output, usage statistics, error handling, and non-streaming response compatibility to ensure comprehensive coverage of scenarios.

TypeScript implementation provides type safety (catching errors at compile time); ecosystem compatibility with mainstream frameworks (Express, Fastify, Next.js, etc.); adherence to the single responsibility principle (focused on streaming protocol conversion, lightweight and easy to test).

Multi-model chat applications (users can switch models without modifying application code); AI agents/complex workflow systems (smooth integration of multi-model interactions); agent orchestration scenarios (tool call recognition and parameter assembly processing).

Handling SSE edge cases (event boundaries, multi-byte characters, connection recovery); compatibility with OpenAI ecosystem differences (subtle deviations in open-source projects like Ollama and vLLM); using the async generator pattern to optimize performance (low memory usage).

llm-stream-assemble addresses real pain points and is worth the attention of developers working on multi-model AI applications. Usage recommendations: Understand the list of supported models and version compatibility; evaluate performance in high-concurrency scenarios; pay attention to project maintenance activity and community feedback. We look forward to more infrastructure projects to promote the engineering of AI development.