sohumsuthar/simulink-mcp

---

14 structured MCP tools for loading models, inspecting blocks, tweaking parameters, running simulations, and capturing figures as images. Every operation is a typed tool call with validated inputs — no arbitrary code execution.

Why this over a general MATLAB MCP server?

General-purpose MATLAB servers pass arbitrary code strings to eval(). That works until the AI hallucinates a function name, forgets a semicolon, or writes a plot() call that blocks forever. This server exposes Simulink operations as discrete, validated tools — the AI picks the right tool and fills in typed parameters instead of generating free-text MATLAB code.

Install

1. MATLAB Engine for Python

The MATLAB Engine ships with every MATLAB installation:

cd /extern/engines/python
pip install .

`` is typically:

• macOS: /Applications/MATLAB_R2024b.app
• Linux: /usr/local/MATLAB/R2024b
• Windows: C:\Program Files\MATLAB\R2024b

2. simulink-mcp

pip install simulink-mcp

Or from source:

git clone https://github.com/sohumsuthar/simulink-mcp.git
cd simulink-mcp
pip install .

3. Verify

python -c "import matlab.engine; print('OK')"

Configuration

Claude Code

claude mcp add simulink -- python -m simulink_mcp

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "simulink": {
      "command": "python",
      "args": ["-m", "simulink_mcp"],
      "env": {
        "SIMULINK_MCP_WORKDIR": "/path/to/your/models"
      }
    }
  }
}

Then ask Claude to load a model and it’ll take it from there.

Environment Variables

Tools

Architecture

Claude    simulink_mcp    MATLAB + Simulink
                              |
                           app.py
                              |
                +------+------+------+---------+
                |      |      |      |         |
            model    inspect  modify  simulate
            (4)      (3)      (5)     (2)

• Lazy engine startup — MATLAB starts on first tool call, not server launch. MCP handshake finishes in ~0.5s; MATLAB boots in the background (~15-20s).
• Stdout isolation — Every MATLAB eval captures stdout/stderr to StringIO. Without this, MATLAB chatter corrupts the JSON-RPC stdio transport.
• Persistent session — The engine survives across tool calls. simulate() stores simOut in the workspace so get_simulation_data() can pull from it later.
• Figure capture — After simulation, open MATLAB figures are exported as PNG, base64-encoded, and returned inline. Figures are closed after capture to prevent accumulation.

Compatibility

simulink-mcp itself requires Python 3.10+ (the upstream mcp package dropped 3.9). Python 3.13 is not yet supported by any MATLAB release. The Engine API has been stable across all listed versions — no code changes needed when upgrading MATLAB.

Troubleshooting

Server times out on startup import matlab.engine is deferred to first tool call by design. If it appears at module level, move it inside the function.

MATLAB engine crashes Restart the MCP server. MATLAB sessions can die on invalid Simulink operations.

PID block “Failed to evaluate mask initialization commands” Common on R2024a when InitialConditionForIntegrator or InitialConditionForFilter falls outside saturation limits. Set both the limit and IC:

set_block_param("model/PID", "LowerSaturationLimit", "0")
set_block_param("model/PID", "InitialConditionForIntegrator", "0")

Contributing

See CONTRIBUTING.md.

License

PolyForm Noncommercial 1.0.0 — free for non-commercial use.

Contact

sohumsuthar.com/contact