QSong-github/DrugClaw

Built on LangGraph, it employs a **Code Agent** that queries heterogeneous biomedical sources in their native formats rather than forcing them through a single abstraction layer, combined with **registry-backed skill organization** that tracks resource availability and metadata in real time. The system supports multiple execution modes—graph-based multi-hop reasoning, simple sequential retrieval, and web-only fallback—enabling precise evidence synthesis for pharmacologically complex questions like drug-target interactions, adverse reaction provenance, and DDI mechanisms. It integrates with OpenAI-compatible LLM endpoints (OpenAI, Azure, vLLM, Ollama) and maintains structured canonical drug name resolution to normalize queries across alias variants before planning and retrieval.