Visual Box is an open-weight, mechanistic interpretability instrument panel for Large Language Models. Built from scratch with a pure NumPy inference engine (with zero PyTorch runtime dependency for SmolLM2) and support for any Hugging Face model, it turns the opaque LLM “black box” into a transparent glass chamber. Users can watch tokens get tokenized, trace attention across 30 transformer layers, inspect next-token probability distributions, project intermediate layer states via Logit Lens, and steer model activations in real time.