Mercury MCP: Cross-Architecture LLM Internal Observation, as Agent Tools

"Most AI coding agents don't know what's inside the model they're talking to. Mercury does."

Mercury MCP exposes a 23-LLM cross-architecture observation database to any agent that speaks the Model Context Protocol (Claude Code, Cursor, Cline, Goose, etc.).

Built entirely on consumer hardware (one Mac mini + one NVIDIA DGX Spark) at near-zero compute cost.

What it answers

Try these prompts with any MCP-aware agent after installing:

• "What hidden dimensions are universally hot across LLM families?"
• "In qwen-7B, which layer has the same functional fingerprint as falcon-7B layer 16?"
• "Compose a layer recipe for a Chinese-writing + reasoning hybrid model."
• "Show me OLMo2's anchor dimensions and how they overlap with the qwen anchor set."

The agent calls Mercury MCP tools; Mercury answers from precomputed observation data.

Why this matters

For mech interp researchers

23-LLM cross-architecture survey across 13 architecture families, two observation tiers per model.

Tier-A (output-layer logit hooks): cheap, fast, candidate-signal screening. May surface artifacts from shared tokenization (vocab-id mod hidden_size collisions). Useful as hypothesis-generating layer.

Tier-B (HF output_hidden_states across all layers): the actual finding layer. Per-layer residual stream fingerprints. Cross-architecture functional layer alignment qwen-7B L15 to falcon-7B L16 reaches 0.868 similarity. 54/84 model-pairs aligned at the middle layers (50-60% depth).

Earlier "dim 11 universal across families" claim from Tier-A is being re-framed as candidate signal, not validated residual stream feature. Working through this in Paper A draft.

For LLM application developers

Stop guessing which model to fine-tune. Mercury tells you structurally which layers carry which capability and which models' middle layers are functionally interchangeable.

For Frankenstein / model-merge people

Cross-architecture per-layer alignment matrix included (12 Tier-B models, 84 pairwise). Strongest single match: qwen-7B L15 to falcon-7B L16 at 0.868 similarity. mistral / yi / gemma occupy alternative residual subspace, possible non-interfering composition substrate.

Install (3 lines)

git clone https://github.com/norika1207-lab/mercury-mcp
cd mercury-mcp
pip install -e .

Then add to your MCP client config:

// ~/.claude/mcp_settings.json  (Claude Code)
{
  "mcpServers": {
    "mercury": {
      "command": "python",
      "args": ["-m", "mercury_mcp.server"]
    }
  }
}

Restart your agent. You now have 7 new tools: mercury_list_models, mercury_anchor_dims, mercury_universal_anchors, mercury_layer_fingerprint, mercury_cross_arch_equivalent, mercury_compose_recipe, mercury_about.

The 23 models in v0.1 (13 architecture families)

Tier-A all 23 done. Tier-B 12 of 23 done, rest in progress.

Findings (still evolving, see Paper A draft for current framing)

1. Cross-architecture per-layer functional alignment. qwen-7B L15 to falcon-7B L16 reaches sim 0.868. 54 of 84 pairwise comparisons show sim above 0.7 at middle layers (50-60% depth). This is the strongest finding right now. Comes from Tier-B (residual stream level), so not vulnerable to tokenizer artifact critique.

2. Distillation appears to transplant anchor structure. DeepSeek-R1:70b uses a llama base but Tier-A anchor structure aligns with qwen at 44.7x random baseline. Note this is Tier-A so caveats apply (see point 4), but the magnitude is hard to dismiss as pure vocab artifact at 44.7x.

3. Three families show alternative residual stream geometry. mistral-7B, mistral-small, yi-1.5 all show 0/11 qwen-anchor presence in Tier-A and structurally different hot dims in Tier-B. They occupy a different subspace, possible substrate for non-interfering cross-architecture composition.

4. Methodology caveat being worked through openly. Tier-A "dim 11 universal across families" (originally framed as residual stream feature) may be a vocab-token-id mod hidden_size collision artifact from shared tokenizer training. Tier-B on OLMo2 contradicts the Tier-A reading. Paper A is being re-framed with Tier-A as screening layer (candidate signals) and Tier-B as finding layer (validated residual stream observations). Discussion welcome.

Full data + analysis: see anchor-survival-MASTER.txt and the paper-handoffs/ repo.

Tool reference

How the data was collected

• Observation protocol: 10 multilingual prompts per model, ~60 tokens each, greedy decoding
• Tier-A: hooked into llama_cpp.Llama.logits_processor to capture output-layer hot dims per token
• Tier-B: transformers.AutoModelForCausalLM.from_pretrained with output_hidden_states=True, per-layer residual stream activation magnitudes binned into 10 quantiles
• Cell grid scheme: addressable (layer, dim, quantile_idx) mmap-able binary, ~24MB-840MB depending on model size
• Source code: paper-handoff paper-G-composable-llm/tools/

Reproducibility: ~4 hours wall-clock per model on a Mac mini M4 Pro. No GPU required.

Cite

@dataset{oda_mercury_2026,
  author = {Chen, Ho Yiing (norika)},
  title  = {Mercury: Cross-Architecture Hot-Dim Geometry Database for 23 LLMs},
  year   = {2026},
  doi    = {10.5281/zenodo.20352085},
  url    = {https://github.com/norika1207-lab/mercury-mcp}
}

License

MIT for code. Data CC-BY-4.0.

Author

norika (Chen Ho Yiing), Taiwan independent researcher.

• ORCID: 0009-0006-6816-9891
• Google Scholar: wrTR3VMAAAAJ
• GitHub: @norika1207-lab

Built solo, no institutional affiliation, no grant, no GPU. Just curiosity and stubborn nights.