Hi HN, We ' ve been frustrated with the fragmentation in the LLM ecosystem. Switching between OpenAI, Anthropic, and Google often means rewriting state management logic or losing model-specific reasoning features. So we built AgentHub to solve this. It’s a small, open-source SDK that provides a unified Python/TypeScript interface for all SOTA models. Why I built this: Existing tools like LangChain felt too heavy, and routers like OpenRouter are closed-source and lack deep execution auditing. It includes a lightweight board for auditing LLM executions. You can permanently trace every run by passing just one parameter. Key Focus: We simplify the engineering process with an asynchronous, stateful, and streaming API specifically designed for multi-turn agentic executions, which significantly flattens the learning curve with zero code changes. And we ensure that model-specific capabilities, such as interleaved thinking and caching, are rigorously validated and aligned across providers with no loss of performance. I’d love to hear your thoughts.