supermemoryai/llm-bridge

Interoperability between input formats of various LLMs, with observability, error handling, etc. built in.

What it solves

Building applications that support multiple LLM providers (like OpenAI, Anthropic, and Google) is difficult because each provider uses different API formats, streaming protocols, and tool-calling structures. LLM Bridge acts as a universal translation layer that normalizes these requests and responses, allowing developers to write code once and deploy it across different AI models without losing data.

How it works

LLM Bridge uses a universal intermediate format to map provider-specific APIs. You can convert an incoming request from any supported provider into this universal format using toUniversal, and then export it to any other provider's format using fromUniversal. The library handles the complexities of mapping multimodal content, tool calls, structured outputs, and streaming events (SSE) between these different systems while ensuring zero data loss through a reconstruction mechanism.

Who it’s for

This library is designed for developers building AI-powered applications, proxies, or middleware who need to support multiple LLM providers simultaneously or switch between them easily without rewriting their core integration logic.

Highlights

  • Perfect Translation: Bidirectional conversion between OpenAI, Anthropic, Google, and OpenAI Responses API formats.
  • Zero Data Loss: Preserves all original fields for perfect request reconstruction.
  • Full Streaming Support: Includes parsers and emitters to handle Server-Sent Events (SSE) across different providers.
  • Advanced Features: Native support for tool calling, multimodal inputs (images/documents), structured outputs, and extended thinking blocks.
  • Type Safety: Built with TypeScript for strict type checking and robust error handling.

Related

  • Project
    Headmaster218/MoE4All

    MoE4All is a high-performance inference engine that enables running large Mixture-of-Experts models on consumer hardware by intelligently offloading weights between SSD, RAM, and GPU VRAM.

  • Project
    rapidsai/dask-cuda

    A set of utilities for deploying and managing Dask workers on CUDA-enabled systems to facilitate GPU-accelerated computing.

  • Project
    HRuiCcc/RuiC-oc-ferry

    OC Ferry is a cross‑platform Electron tray app that automatically discovers free OpenCode models, validates they can follow WorkBuddy’s tool‑calling JSON envelope, caches the results, and writes compatible entries into WorkBuddy’s `models.json`. It also runs a local OpenAI‑compatible endpoint (127.0.0.1:41980) that forwards WorkBuddy requests to those models. No login or API key is needed; the app bundles an isolated OpenCode runtime and handles concurrent probing, caching, and safe import/export of model entries.

  • Project
    flagos-ai/FlagAttention

    FlagAttention is a Triton‑based library offering fast, memory‑efficient attention kernels (FlashAttention, Piecewise, paged, sparse, quantised, recurrent, etc.) for training and inference of large models across many accelerators. Stable dense operators are ready for production; alpha variants provide experimental features like KV‑cache paging, block‑sparse decoding, and INT8 quantisation.