norika1207-lab

norika1207-lab / mercury-mcp

Public

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

github.comnorika1207-labmercury-mcp ai-agents anchor-dimensions consumer-hardware cross-architecture frankenstein-merge
13
2
89% credibility
Found May 24, 2026 at 25 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Mercury MCP is a research database that exposes pre-computed observations about how 23 different AI language models work internally. It connects to AI coding assistants and lets users query which internal dimensions are most active, find similar layers across different model architectures, and compose custom hybrid model recipes. Built by an independent researcher on consumer hardware, it includes findings like cross-architecture layer similarities (86.8% match between Qwen and Falcon layers) and identifies which model families share structural patterns. The project is openly documented with academic citations and a Zenodo DOI.

How It Works

1
🔬 You discover a way to see inside AI models

You learn about a research database that maps the internal structure of 23 different AI models, built by an independent researcher on consumer hardware.

2
🧩 You connect it to your AI assistant

You add Mercury to your AI coding assistant's configuration so it can access the database when you ask questions.

3
📋 You explore what's available

Your assistant shows you all 23 observed models across 13 different architecture families, from Qwen to Mistral to OLMo2.

4
🔥 You peek at a model's hot spots

You ask which internal dimensions consistently fire across different prompts in a specific model, and your assistant reveals the patterns.

5
You choose your path
🔍
Find similar layers across models

Discover that layer 15 in one model works almost identically to layer 16 in another (86.8% similarity)

🧪
Build a hybrid model recipe

Request a combination of Chinese writing, reasoning, and code capabilities mapped to specific layers

✨ You gain structural insights

You understand which layers carry which capabilities across models, enabling smarter decisions about fine-tuning or model merging.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 25 to 13 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is mercury-mcp?

Mercury MCP is a Python-based MCP server that gives AI coding agents x-ray vision into the LLMs they're working with. It exposes a precomputed database of internal observations from 23 LLMs across 13 architecture families as 7 queryable tools. Instead of guessing which layers handle which capabilities, agents can ask questions like "which layer in qwen-7B matches falcon-7B layer 16?" and get data-backed answers. The tools let you list models, inspect anchor dimensions, find cross-architecture equivalents, and even compose layer recipes for model merge experiments.

Why is it gaining traction?

The hook is clear: mechanistic interpretability meets agent tooling. Most developers have no idea what happens inside the models they deploy. Mercury flips this by making internal geometry queryable through the same Model Context Protocol that Claude Code, Cursor, and Cline already speak. The cross-architecture layer alignment data is the killer feature—qwen-7B layer 15 maps to falcon-7B layer 16 at 0.868 similarity, and 54 of 84 model pairs show strong middle-layer alignment. This is the kind of data that makes model merging less guesswork and more engineering.

Who should use this?

Mech interp researchers doing cross-architecture surveys will find the most value here. LLM application developers tired of guessing which model to fine-tune can use the anchor and fingerprint tools to make data-driven decisions. If you're into Frankenstein-style model merging or layer-swap experiments, the cross-architecture equivalent finder is purpose-built for you. Regular developers using AI coding agents probably don't need this yet—unless you're building tooling that works with multiple LLM families.

Verdict

Mercury MCP scores 0.899% on credibility—the author is a solo researcher with an ORCID, Zenodo DOI, and transparent methodology. With only 13 stars the community is nascent, but the README is thorough, the data is reproducible on consumer hardware, and the honest framing of methodology caveats builds trust. This is a niche tool for a niche problem, but if you're doing cross-architecture LLM research or model merging, the foundation is solid and the approach is sound. Watch this one.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.