Your architecture, opened up.
Open a PyTorch class from VS Code. Explore its layers and nested modules in a local browser.
FOR PEOPLE WHO BUILD WITH PYTORCH
See your model.
Understand the change.
Explore its architecture, follow the shapes,
and review the Python behind every change.
Local VS Code extension · Your model runs on your machine
ONE MODEL. SIX WAYS TO UNDERSTAND IT.
Interactive example · No setup
ResidualClassifier
Input[1, 64]→StemLinear · 64 → 128→ResidualBlock4 operations inside→HeadLinear · 64 → 10→OutputModel result
SelectedModuleRecorded failureChoose a layer to look inside
A real model. A recorded run. Your turn to explore.
Browse the model as a list
THE ACTUAL WORKFLOW
From your file to a clearer view.
FROM SOURCE TO UNDERSTANDING
Keep the code.
Get the whole picture.
TensorViz connects the way you write a model
with the way you reason about it.
Find the layer behind the error.
Run validation when you choose. Inspect the observed shapes and the source behind a failure.
A change you can see.
Propose an edit, review the native Python diff, then choose Apply or Discard in VS Code.
BRING YOUR OWN MODEL
Make the next
model change clearer.
Start with a local PyTorch project in VS Code.
Get started with TensorViz Early preview · macOS Apple Silicon verified
A few useful details.
Where does TensorViz run?
The extension runs in local VS Code. Your graph opens in a local browser, and Python execution uses your selected environment on your machine. The landing-page example replays previously captured evidence.
Does opening a graph run my model?
No. Opening and refreshing the architecture parses source, including unsaved buffers. Validation is an explicit action on saved source. Trace on save is optional and off by default.
Can I use it with my coding agent?
Compatible external clients can connect through TensorViz’s MCP endpoint. They can inspect the model and propose changes for native review. TensorViz has no built-in chatbot and does not ask for provider keys.
What can I share?
A portable .tensorviz.json graph with saved views,
walkthroughs and notes. Recipients can explore it in the static
viewer without Python or the extension. Source code and weights
are not included.