The repository describes a local interface that maps model layers in 3D, with one point per neuron, and lets users inspect a neuron's steering and logit lenses.
The tool can load Qwen3.5-4B locally on MPS, CUDA, or CPU for steered chat, activation inspection, next-token log probabilities, and ablations. Its sweep tooling pushes MLP neurons up and down across prompts, retains rollout outputs, and records per-neuron statistics. The supplied explorer caches are included in the repository. The documented large sweep setup uses a CUDA 13 GPU pod and a steering fork of vLLM.
The interface also includes a Claude CLI integration exposed through an MCP server. According to the repository, that integration can query neuron data, run steering and ablations, and control views in the local explorer. The source does not establish that the visual groupings correspond to robust model mechanisms, that steering effects generalize beyond the tested prompts and model, or that the tool has been independently evaluated.