The model-first environment.

03See

Look inside the model, not across files.

One forward pass and a PyTorch model draws itself: modules, operations and measured shapes, with per-module latency when you ask for it. Every node is one click from the Python that defines it, and your cursor leads back to the node.

• PyTorch · measured shapes
• JAX and Flax · open as traced graphs

Illustration: a transformer block unfurls from its input as a graph of node cards, with tensor shapes ticking in along the edges, from [1, 128, 768] to [1, 128, 3072] and back. Selecting the attention node, blocks.3.attn, lights its exact line in model.py, line 18. A cursor on the mlp line in the code lights the MLP group in the graph.

04Change

Edit the graph. Ship the diff.

Change a parameter, swap an activation, add a layer or delete one on the graph, and tensorViz writes it as a native code diff that you apply or discard. Any Git change or GitHub pull request opens as a graph diff beside the code diff, without running either version.

Reviewers see what the model became, not just which lines moved.

Illustration: on the graph, the MLP’s GELU activation is replaced with SiLU from the node’s menu. The GELU node fades out in red, the SiLU node grows in green, and the edges downstream flash amber while their shapes hold at [1, 128, 3072] and [1, 128, 768]. Beside it, model.py shows the matching hunk, minus nn.GELU() and plus nn.SiLU(), with shapes unchanged and Apply and Discard. The same line changes all 12 blocks.

05Humans and agents

Agents now write model code. Give them the model, not the files.

They change it faster than anyone can read the diff, and they work from files and guesses. In tensorViz they work on the same representation you do: they call get_graph, inspect_nodes and compare_runs to read the model’s actual graph and shapes, and through the VS Code extension they call propose_edit. Every proposal comes back as a draft graph beside the code diff, and you accept or discard it.

One-click MCP setup for Claude Code, Codex and Cursor from the VS Code extension, and any MCP client by hand. Your agent brings its own model, so no tensorViz key is needed.

Illustration: two cursors, you and an agent, on the same graph. The agent’s tool calls stream in: get_graph, then inspect_nodes on blocks.3.attn, then propose_edit with 16 heads instead of 12. Within a second the proposal appears as a dashed amber draft beside the attention node, and model.py shows the matching diff, heads=12 to heads=16, with shapes unchanged and Accept and Discard. You accept, and the draft merges into the graph, the code and all 12 blocks.