mcpserver.lol
registry/neurarch-mcp
Connection check local install

neurarch-mcp

Reads a PyTorch .py, a Neurarch graph, or a HF repo; agents inspect, lint, verify and rank designs.

Tools 19
GitHub stars 1
Installs / wk
Licence MIT
Transport stdio
Last checked never

Tools & capabilities

19 tools

Read from the published package source — this server runs locally, so there is no endpoint to query. Names are taken from the code, not observed at runtime, and descriptions are often absent.

add_connection from source
architecture from source
A library id from list_architectures, e.g.
compare_with_reference from source
Put the current model next to a published one from the bundled library and explain the structural differences that matter.
delete_connection from source
explain_finding from source
What a lint rule or check_design finding means for this model, the evidence behind it, and the smallest edit that clears it.
flops_by_block from source
focus from source
Optional: what to review for, e.g.
get_layer from source
get_model_summary from source
list_blocks from source
list_connections from source
list_hyperparams from source
mermaid_diagram from source
param_count_by_block from source
pre_train_checklist from source
The checks worth running before spending GPU time: structural, design rules, cost, GPU fit, and what is still unknown about the graph.
review_design from source
A structured design review of the current model: readiness, risks, parameter budget, and the edits worth making, every number from the tools.
rule from source
The rule id or finding title, e.g.
shrink_for_target from source
Find the edits that bring the model under a parameter, memory, latency or GPU budget with the least damage, and rank the resulting variants.
target from source
The budget, e.g.