Launching today
NeuronScope
Traces LLM outputs to the neurons and heads that caused them
2 followers
Traces LLM outputs to the neurons and heads that caused them
2 followers
NeuronScope is a CLI and MCP server for mechanistic interpretability. It is built on TransformerLens. Point it at an open-weight model to instantly see which attention heads and neurons drive predictions. Supporting hundreds of checkpoints, it avoids strict allowlists. Its native MCP server and versioned JSON let AI agents (Claude, Cursor) directly trace, ablate, and sketch circuits without any Python coding requirement.
When debugging LLMs, seeing the output isn't enough. Understanding the why is the real bottleneck. "It's probably the prompt" is not a scalable debugging strategy for specific hallucinations.
We built NeuronScope. It wraps TransformerLens into a stable CLI and MCP server, allowing scripts or agents to ask, "Which attention heads and neurons drove this token?" and instantly receive versioned JSON instead of digging through raw tensors.
Core Commands:
- Trace: Rank heads/neurons by their contribution to a predicted token.
- Activations: Get full per-layer activation stats.
- Patch: Zero-ablate components and observe the delta.
- Circuit: Build a best-effort circuit sketch.
It supports all 249 TransformerLens checkpoints (Llama, Qwen, Pythia, etc.). The tool is fully MIT-licensed and operates entirely locally, requiring no accounts or API keys.
Repo: https://github.com/RudrenduPaul/NeuronScope
MCP Servers:
https://mcpservers.org/servers/rudrendupaul/neuronscope
https://glama.ai/mcp/servers/RudrenduPaul/NeuronScope
PyPI: https://pypi.org/project/neuronscope-cli
Get started:
pip install "neuronscope-cli[mcp]"
Thoughts and feedback from the community are very welcome, especially where the circuit heuristic hits its limits.