norika1207-lab/mercury-mcp

View on GitHub ↗

Cross-architecture LLM internal observation database (23 models, 13 architecture families). Exposed as MCP tools for any AI coding agent.

43 ★14 forksPythonUpdated 3mo ago

What you need to know

MCP server exposing a cross-architecture LLM internal observation database (23 models, 13 architecture families) as agent tools for mechanistic interpretability.

Install

git clone && pip install -e .

Key features

  • 7 tools: list_models, anchor_dims, universal_anchors, layer_fingerprint, cross_arch_equivalent, compose_recipe, about
  • Tier-A output-layer logit hooks and Tier-B residual stream fingerprints
  • Cross-architecture layer alignment data
  • Built on consumer hardware (Mac mini M4 Pro)

Best for

Mechanistic interpretability researchers and model-merge practitioners.

Platforms: Python

Reviewed 2026-08-11

Topics

ai-agentsanchor-dimensionsconsumer-hardwarecross-architecturefrankenstein-mergellmmcpmechanistic-interpretabilitymodel-context-protocolopen-datatransformer
Stars
43★
Forks
14
Language
Python
License
Created
2026-05-23
Last push
2026-05-24