Launching today

tensorViz
Let your coding agent see inside your PyTorch model
68 followers
Let your coding agent see inside your PyTorch model
68 followers
Coding agents can read your PyTorch code, but they never see what actually flows through the model. tensorViz runs one real forward pass and turns it into an interactive graph with every module and tensor shape. Claude Code, Codex and Cursor can search that graph and read the measured shapes, so they fix model code from evidence instead of guesses. Or skip the install: the Model Explorer has graphs of 560+ models from recent papers to browse and discuss. Free and open source (MIT).







Free
Launch Team

LoomTalk it through in Loom. Ship with AI
Promoted
Hey Product Hunt ๐ I'm Ayush, and we build tensorViz.
If you've asked Claude Code or Codex to change a PyTorch model, you've probably seen this: the agent edits one layer, and three modules later there's a shape mismatch it can only guess about. Agents read the source, but they never see what actually flows through the model.
tensorViz gives your agent, and you, the real picture:
Runs one forward pass of your local nn.Module, with example inputs you save once
Shows every module and tensor shape as an interactive graph, beside your agent's chat or in a local browser tab
Click a node and your agent knows exactly what you selected
Your agent can search the graph for any layer, ops or parameter and read its measured shapes
Drill from the whole model down to a single layer, on a light or dark canvas
No install needed to look around: the Model Explorer. At tensorviz.dev/explore we've published graphs of 560+ models from recent papers, with links to each paper and its code. Filter by area (language, vision, video, robotics, speech and more), size or release date, open any graph in your browser and start a discussion right beside its layers.
Stuck on your own model? Add its graph from our plugins or extension, keep it private or share it with teammates or by link, and tag the exact layers you're asking about. Graphs hold structure and shapes, never your code or weights.
Free, open source (MIT) and local. Your model runs on your machine, with no account and no API key. You only sign in to join a discussion or share a graph.
Honest limits for this preview:
CPU forward passes; GPU and training runs aren't covered yet
Very large models can produce a partial graph
๐ Launch-day offer: drop a link to a public PyTorch repo below and I'll reply with its graph.
And I'd love to know: which coding agent do you use for ML work? I'll be here all day.
I use Claude Code, and this is a failure mode I recognize โ the agent reads the code fine, then guesses at what actually flows through it and loops on shape errors. Grounding it in one real forward pass instead of the static code is the right call. For the very large models that only get a partial graph, does the agent know which parts are missing, or can it mistake the partial graph for the whole model?
@jcbuildsย Yes, Claude can tell when a graph is partial and see which modules the capture missed. We track coverage at the module level, so the individual operations and tensor shapes inside those gaps stay unknown.
Shape mismatches are exactly where I've seen agents go in circles. Nice idea.
Shape mismatches three modules downstream are such a pain to debug, so grounding the agent in a real forward pass makes sense. One thing I'd wonder about is models with dynamic control flow or variable sequence lengths. Does the graph reflect just the example input you saved, or can it show how shapes change across a few different inputs?
@galdayanย It changes dynamically. The code is open source: https://github.com/tensorviz-org/tensorviz