neurarch-mcp

Structural awareness of a PyTorch model for the agent. Point it at a `.py`, a `.neurarch.json`, one of 81 bundled reference architectures (`zoo:<id>`) or a Hugging Face repo (`hf:<org/name>`, behind `--hf`). **The ladder:** `validate_model` (is it a graph) -> `lint_model` (design rules, with provenance) -> `check_design` (readiness, cost, deployment fit) -> `rank_designs` (which of k candidates deserves the training run; blocked last, ties stay ties, calibration in every result). Plus `describe_architecture`, `layer_impact` (blast radius), `export_pytorch`, `list_architectures`, `find_models`, five prompts (`/review_design`, `/pre_train_checklist`, `/shrink_for_target`, `/compare_with_reference`, `/explain_finding`) and resources for the graph, its Mermaid diagram and generated source. Offline: the parser, rule engine, verifier, ranker and reference library are vendored. No API key, no network call unless you turn one on. Also a CLI: `npx neurarch-mcp lint model.py`. - npm: https://www.npmjs.com/package/neurarch-mcp - Source: https://github.com/neurarch-ai/neurarch-mcp - Docs: https://www.neurarch.com/mcp - Real shapes from real code: `pip install neurarch-trace`

First seen 2 Oct 2026. Evidence as of 12 Oct 2026.