@infatoshi/gpu-mcp-server

@infatoshi/gpu-mcp-server

Pure Node.js MCP server for compiling and running CUDA code on Modal's cloud GPUs. No Python dependencies - just install via NPM and run CUDA code on B200/H100 GPUs directly from Cursor or any MCP-compatible editor.

Features

• Zero Python dependency - Pure Node.js implementation
• Smart defaults - Auto-detects GPU architecture, entry files, and compilation flags
• One command install - npm install -g @infatoshi/gpu-mcp-server
• Minimal config - Only requires files_glob parameter
• Auto-deployment - Deploys Modal app automatically on first use
• Fast iteration - Modal caching speeds up repeated compilations

Prerequisites

• Modal account with GPU access (sign up)
• Modal CLI authenticated:

  pip install modal
  modal token new
  

Installation

npm install -g @infatoshi/gpu-mcp-server

Configuration

Add to your ~/.cursor/mcp.json or VS Code MCP config:

{
  "mcpServers": {
    "gpu-cuda": {
      "command": "gpu-mcp",
      "args": []
    }
  }
}

That's it! No environment variables needed - the server uses smart defaults.

Usage

Basic Usage

In Cursor, just ask the AI to compile your CUDA code:

"Compile and run vector_add.cu"

The MCP tool only requires one parameter:

• files_glob: Pattern to match CUDA files (e.g., "*.cu", "src/**/*.cu")

Advanced Usage

Optional parameters for fine-tuned control:

{
  "files_glob": "kernels/*.cu",
  "project_root": "/path/to/project",  // Auto-detected if omitted
  "entry": "main.cu",                  // Auto-detected if omitted
  "arch": "sm_90",                     // Auto-detected from MODAL_GPU if omitted
  "extra_flags": ["-O3", "--use_fast_math"],
  "run_args": ["1000000"]
}

Environment Variables (Optional)

# GPU type (default: B200)
export MODAL_GPU=B200  # or H100, H200, A100

# Modal app name (default: cuda-kernel-runner)
export MODAL_APP=my-cuda-app

GPU Support

GPU Type	Architecture	Compute Capability
B200 (default)	sm_100	10.0
H100/H200	sm_90	9.0
A100	sm_80	8.0

How It Works

• Smart Detection: Auto-detects entry file, GPU architecture, and compilation flags
• File Collection: Gathers all CUDA files matching your glob pattern
• Modal Deployment: Deploys the Modal app if not already deployed
• GPU Compilation: Compiles on Modal's cloud GPUs with nvcc
• Execution: Runs the compiled binary and returns output

Example

Create vector_add.cu:

#include <stdio.h>

__global__ void vectorAdd(float *a, float *b, float *c, int n) {
    int i = blockDim.x * blockIdx.x + threadIdx.x;
    if (i < n) c[i] = a[i] + b[i];
}

int main() {
    const int N = 1000000;
    // ... CUDA setup code ...
    vectorAdd<<<(N+255)/256, 256>>>(d_a, d_b, d_c, N);
    printf("Vector addition successful!\n");
    return 0;
}

In Cursor:

"Run vector_add.cu on GPU"

The server automatically:

• Detects B200 GPU (sm_100 architecture)
• Uses vector_add.cu as entry file
• Compiles with -O3 --use_fast_math
• Runs on Modal's B200 GPU
• Returns compilation and execution logs

Troubleshooting

"Not authenticated with Modal"

modal token new

"Function not found"

• The server will auto-deploy on first use
• Check modal app list to verify deployment

Want to see what's happening?

• Check stderr output for compilation and deployment logs
• Modal dashboard shows function invocations

License

MIT

Author

infatoshi