LiteLLM seemed to be the best approach for what I needed - simple integration with different models (mainly OpenAI and the various Bedrock models) and the ability to track costs / limit spending. It's working really well so far.

Didn't know about LiteLLM. That seems to be the right kind of middleware most people would need, instead of Langchain.