TokenSage sits between your application and the model, removing unnecessary conversation context before it reaches the provider. The optimizer itself uses no LLM.