MCP integration

MCP integration#

VisualTorch provides an optional Model Context Protocol (MCP) server for generating architecture diagrams from model source supplied by an MCP client. The MCP server uses VisualTorch’s public visualtorch.render(...) API and writes the generated image to a local file.

Installation#

Install the optional MCP dependency alongside VisualTorch:

python -m pip install "visualtorch[mcp]"

For a source checkout, install the project in editable mode:

python -m pip install -e ".[mcp]"

Both commands install the visualtorch-mcp console script.

MCP client configuration#

Add the server to an MCP client configuration that supports stdio servers:

{
  "mcpServers": {
    "visualtorch": {
      "command": "visualtorch-mcp"
    }
  }
}

If the client should use a specific Python environment, configure its interpreter explicitly:

{
  "mcpServers": {
    "visualtorch": {
      "command": "C:/path/to/venv/Scripts/python.exe",
      "args": ["-m", "visualtorch_mcp.server"]
    }
  }
}

Tools and resources#

The visualize_model tool accepts:

  • source: Python source that imports its dependencies and defines the model.

  • input_shape: one input shape including the batch dimension, or one shape per positional input.

  • style: graph, flow, or lenet. The compatibility aliases layered, layered_view, lenet_style, and lenet_view are also accepted.

  • model_expression: an expression evaluated after source; it defaults to model and can be set to an expression such as Net() or build_model().

  • output_path: an optional image path. Relative paths are resolved under output_dir.

  • output_dir: an optional base directory for generated images.

  • options: VisualTorch render options, such as {"palette": "dracula", "show_dimension": true}.

  • workdir: an optional working directory for the model source.

  • timeout_seconds: the render subprocess timeout, defaulting to 120 seconds.

The visualtorch_reference tool and the visualtorch://docs resource expose links to the upstream VisualTorch API references and examples. The visualtorch://version resource reports the MCP integration version.

Security#

The server intentionally executes the supplied model source and evaluates model_expression in a separate Python subprocess. Only connect it to clients and use source code that you trust. The subprocess timeout and boundary keep a failed render from crashing the MCP server, but they are not a security sandbox for hostile Python code.

Contents