Mercury MCP
An MCP server exposing a precomputed observation database of 23 large language models across 13 architecture families, letting a coding agent ask which internal layers of different models behave alike, without running any of those models itself.
Mercury MCP answers questions such as which hidden dimensions are consistently active across model families, which layer of one model has the same functional fingerprint as a layer of another, or how to compose a layer recipe for a hybrid model, drawing on precomputed observation data rather than live inference. Its data has two tiers: a cheap, fast output-layer logit-hook tier used for candidate screening that can surface artifacts from shared tokenization, and a slower tier that hooks hidden states across every layer, treated as the tier that actually validates a finding, including a per-layer residual-stream alignment matrix across 84 model pairs.
The project's stated audiences are mechanistic-interpretability researchers doing a cross-architecture survey, application developers deciding which layers of a model carry which capability instead of guessing where to fine-tune, and people merging models who want to know which models' middle layers are functionally interchangeable and which occupy a different, potentially non-interfering subspace. It documents at least one earlier finding from the faster tier being walked back to a candidate signal rather than a validated feature once tested against the slower tier.
It was built on consumer hardware, a Mac mini and an NVIDIA DGX Spark, is written in Python, and installs with pip.