# tensorViz coding-agent plugin

The plugin lets Claude Code, Codex, or Cursor open a local PyTorch model, capture one forward call, and read the graph. It shows the same canvas as the editor extension. You do not need the extension to use the plugin.

![Annotated model preview in the shared Canvas, also used by the coding-agent plugin](/images/explore-graph.png)

Use one TensorViz host per agent. Do not connect the same agent to both the plugin and the extension.

## Install

Use your client’s terminal CLI. If it is not installed, start with the [installation requirements](/install.html#plugin).

The published plugin is a preview for Apple Silicon macOS with CPython 3.12 and CPU capture. It downloads a verified compiled runtime on first connection, so you do not need a separate Node.js installation. Allow internet access for that first download.

Codex:

```sh
codex plugin marketplace add tensorviz-org/tensorviz-plugins
codex plugin add tensorviz@tensorviz-standalone
```

Claude Code:

```sh
claude plugin marketplace add tensorviz-org/tensorviz-plugins
claude plugin install tensorviz@tensorviz-standalone --scope user
```

Cursor: follow [Install in Cursor](https://github.com/tensorviz-org/tensorviz-plugins#install-in-cursor).

For a local source archive, see [Codex and Cursor local installation](/reference/agent-plugin.html#install).
Non-interactive Codex installation needs
`sh /path/to/tensorviz/plugins/tensorviz/scripts/install.sh --apply --host plugin`.
To keep a current connection, use `--apply --host existing --connection NAME`
with a name from the installer preflight.

T3 Code uses your Codex profile. See the [public catalog](https://github.com/tensorviz-org/tensorviz-plugins).

Start a new chat afterward. Platform limits and source builds are on the [Install](/install.html) page.

## First graph

1. Open your model project in the client and start a new chat.
2. Ask, for example: "Open `Net` in `models/net.py` as a TensorViz graph."
3. When asked, choose a CPython 3.12 environment that can import the model, and confirm example inputs.

Opening a graph executes one forward call. Ask to **inspect source only** if you want a preview that does not run the model. Saved `tensorviz.toml` recipes are reused. After you edit the source, ask the agent to refresh the graph.

## Model graph panel

Ask the agent to show the **Model graph** panel. Clients without a panel can open the graph in a local browser page instead. An empty panel does not start Python.

Closing the panel does not cancel a capture or edit source. Reopen **Model graph** if it disconnects; the capture is kept.

## Publish to tensorviz.dev

**Publish**, in the panel and browser graph toolbars, adds the open graph to your
[tensorviz.dev](https://tensorviz.dev/explore/) account as private, as the editor
extension does. The first time, choose **Continue in browser** to sign in with Google.
Then open the graph's Share dialog to choose who can see it. Publish is a button in
the graph, not an agent tool. To sign out, run `node scripts/launch.cjs --sign-out`
from the installed plugin directory. Details: [Add graphs to tensorViz](/share-to-website.html#publish-from-the-coding-agent-plugin).

## Permissions

On first use, TensorViz asks for permission to run model and setup code, save `tensorviz.toml`, and install reviewed dependencies in the selected environment. That permission applies to your OS user across projects; your client’s tool controls still apply.

- Model code runs with your user permissions.
- Graph text returned to the agent can become part of that client's conversation.
- Consent does not start a model on its own. Unknown inputs and interpreter choices still need an answer.
- Uninstalling the plugin does not delete the saved permission.

Disable the plugin in your client to stop that client from using its tools. Source-build status and revocation commands are in the [source-package reference](/reference/agent-plugin.html#permissions).

## Help

| Problem | What to try |
| --- | --- |
| No tensorViz tools in a chat | Check that the plugin is enabled, then start a new chat |
| Duplicate tools | Disable the older TensorViz plugin or MCP entry. Keep one host |
| Python environment rejected | Select CPython 3.12 for this preview |
| Capture succeeds but no graph appears | Ask the agent to show the Model graph panel or open the browser graph |

More: [environment recovery](/environment-recovery.html), [setup and inputs](/graph-setup.html), [support and compatibility](https://github.com/tensorviz-org/tensorviz-plugins/blob/main/SUPPORT.md).

## Develop the plugin

Building from source, the source archive installer, service details, and panel limits are in the [plugin reference](/reference/agent-plugin.html) and [plugin build](/plugin-build.html).