DVAI-Bridge

Appearance

DVAI-BridgeLocal AI in any app.

One SDK on every platform. The model runs on the user's device. Any agent framework that speaks OpenAI works without changes.

Get started
View on GitHub
DVAI-Bridge Banner

One import, every platform

Swift on iOS, Kotlin on Android, Dart on Flutter, TypeScript on the web and React Native, C# on .NET — same OpenAI HTTP wire on each.

The model runs on the device

No cloud calls. No per-token bill. No data leaves the user's hardware. Works offline.

The engine picks itself

Apple Foundation Models, MLX, llama.cpp, CoreML, MediaPipe, LiteRT — selected per device. Override per request when you need to.

Your agents already work

LangChain, the OpenAI SDKs, Vercel AI SDK, CrewAI, LlamaIndex — point them at the local URL and run.

Multi-modal

Text, images, audio. Gemma 4, LLaVA, Whisper. Runs in a worker; the main thread stays free.

Built for production

Auto-recovery on bad output, generation timeout, bounded reload — users see a brief reload, not a broken screen.

Pair a phone with a desktop

DVAI Hub offloads heavy inference from a phone to a paired laptop on the same Wi-Fi. End-to-end encrypted.

Six SDKs, five registries

npm, Maven Central, CocoaPods, pub.dev, NuGet. Install the one for your stack — the rest are the same shape.

Embedded in your app

Not a gateway. Not an end-user-installed server. Ships inside the app you build.

Why DVAI-Bridge? ​

DVAI-Bridge embeds an OpenAI-compatible AI server inside your app — so the same agent code you wrote for cloud OpenAI keeps working, just locally.

The model runs on the user's device. The HTTP server runs in your app's own process. No cloud account. No per-token bill. No install for your user.

It is not a gateway like LiteLLM. It is not a server your user has to install like Ollama. It is not a framework SDK like LangChain. It is the thin layer between your code and the model — the same shape on iOS, Android, Flutter, React Native, .NET, and the web.

Capabilities ​

  • Text generation — LLM chat and completions on every platform.
  • Embeddings — Vector embeddings for RAG and semantic search.
  • Multi-modal — Text + image + audio via Gemma 4, LLaVA, Whisper, and more.
  • Streaming — Server-Sent Events for token-by-token output, the same way OpenAI streams.
  • Offload to a paired device — DVAI Hub lets a phone hand heavy inference to a laptop on the same Wi-Fi.
  • Auto-recovery — Detects bad output, restarts the engine, retries the request — silently.

System overview ​

Three layers, the same on every platform.

  • The engine — llama.cpp, Apple Foundation Models, MLX, CoreML, MediaPipe LLM, LiteRT, WebLLM, or Transformers.js. Picked at runtime.
  • The HTTP server — 127.0.0.1:38883, /v1/chat/completions, /v1/embeddings, /v1/models. The OpenAI wire, exactly.
  • The SDK — one idiomatic surface per language. Same method names, same lifecycle, same behaviour.

Next steps ​