Moving to an OpenAI-Compatible API in an Afternoon

The most expensive part of switching AI vendors is rarely the models. It is the integration: new SDKs, new auth flows, new response shapes, weeks of testing. OpenAI-compatible APIs exist to make that cost disappear, because the wire format is the same regardless of which model answers underneath.

Migration becomes a configuration change. Point your client at a new base URL, set an API key, and existing code that speaks OpenAI's chat completions or embeddings format keeps working. The same SDKs, the same request bodies, the same response parsing.

That property matters most for teams with many services. A monorepo with fifteen microservices calling three vendors spends weeks on each quarterly model swap. Behind an OpenAI-compatible gateway, the swap is one pull request that edits environment variables.

Watch for the edges, not the happy path. Streaming responses, tool calls, and structured outputs have vendor-specific quirks even when the base format matches. Test those paths explicitly during migration, and keep a compatibility test suite that runs against the gateway rather than any single provider.

llm-kita exposes a single OpenAI-compatible endpoint for video, image, TTS, STT, and presentations. Change the base URL, and SDKs you already have start working against all five modalities.