Mercury MCP v0.1.0 — Cross-Architecture LLM Internal Observation Database for 23 LLMs

Mercury MCP is a Model Context Protocol server that exposes a 23-LLM cross-architecture observation database to any AI coding agent (Claude Code, Cursor, Cline, Goose).

Built entirely on consumer hardware (one Mac mini M4 Pro + one NVIDIA DGX Spark). Total compute cost approximately $0.

Headline findings:

• Cross-architecture residual-stream layer alignment: qwen-7B layer 15 ↔ falcon-7B layer 16 at fingerprint similarity 0.868
• Middle layers (50-60% depth) most universally aligned across families (33/51 model-pairs at sim ≥ 0.7)
• DeepSeek-R1:70b inherits qwen anchor structure through distillation — opens model lineage forensics
• Three families (mistral, mistral-small, yi) show alternative residual-stream geometry — substrate for cross-architecture non-interfering composition

Tools provided: mercury_list_models, mercury_anchor_dims, mercury_universal_anchors, mercury_layer_fingerprint, mercury_cross_arch_equivalent, mercury_compose_recipe, mercury_about

Models surveyed: qwen2.5 (3B/7B/14B/32B/coder-32B), DeepSeek-R1 (7B/32B/70B), llama (3.1-8B/3.3-70B), phi3 (14B), falcon3 (7B), internlm2 (7B), mistral (7B-v0.3/small-3.2-24B), AllenAI OLMo2 (7B), gemma2 (9B), IBM Granite (8B), yi-1.5 (9B), Cohere command-r (35B), starcoder2 (15B), codestral (22B).