A local switchboard for 300+ LLMs that refuses to be a proxy

What it does Adaline Gateway is an isomorphic TypeScript SDK that normalizes chat, streaming, and embedding requests across over 300 LLM providers. It runs fully local—no proxy server required—and ships with optional per-provider packages so you only install what you need. Under the hood it handles retries, batching, caching hooks, and OpenTelemetry instrumentation.

The interesting bit Instead of thin HTTP wrappers, Gateway introduces its own typed prompt abstractions (MessageType, ConfigType, ContentType) and silently translates them to each provider’s quirks. An “assistant” role becomes “model” for Gemini; unknown config fields evaporate instead of erroring. It is fundamentally well-engineered glue code, but the emphasis on local execution and programmatic schema introspection makes it more than a typical fetch wrapper.

Key highlights

• Optional provider packages keep bundle weight down
• Exposes per-model schemas programmatically (max tokens, supported roles, modalities)
• Plug-in system for custom providers, cache backends, and HTTP clients
• Built-in batching, exponential backoff retries, and callback hooks
• OpenTelemetry integration for existing observability stacks

Caveats

• Embedding support is uneven; the provider table shows several major vendors (Anthropic, Google AI Studio, Groq, Bedrock) still lack it
• ConfigType deliberately ignores unknown fields, which can silently swallow typos
• Enterprise trust claims in the README are not backed by named customers or case studies

Verdict Worth evaluating if you’re orchestrating multiple LLM providers in TypeScript and want normalization without the operational overhead of a proxy. Probably overkill if your stack is married to a single provider and you don’t need the type translation layer.