@infatoshi/b200-mcp

Model Context Protocol server that launches a Modal B200 GPU, uploads a single Python script, executes it, and returns stdout/stderr to the MCP client. Designed for npx execution so editors like Cursor can run it without prior installation.

Quick start

{
  "mcpServers": {
    "b200-modal": {
      "command": "npx",
      "args": ["-y", "@infatoshi/b200-mcp"]
    }
  }
}

Invoke the run_python tool with the relative path to the script in your workspace. The server resolves the script, deploys a Modal function if needed, and runs the file on a single B200 GPU.

Tool parameters

| name | type | required | description | | --- | --- | --- | --- | | script_path | string | yes | Relative path to the Python file to execute. | | project_root | string | no | Base directory to resolve script_path. Defaults to the CWD where the MCP server is launched. | | args | string[] | no | Positional arguments forwarded to the Python process. | | env | record | no | Extra environment variables for the script. |

Configuration

Runtime defaults live in config/modal.yaml. Adjust GPU, image, or timeout there instead of hardcoding values. The server reloads the YAML on each invocation so changes take effect immediately.

Modal prerequisites

1. Install the Modal CLI (pip install modal).
2. Authenticate once: modal token new.
3. Ensure your account has access to B200 GPUs.

Development

npm install
node src/server.js

The MCP server uses ESM modules, keeps I/O in adapter/, and isolates pure helpers under lib/.