LoganChu/duke-dcc-mcp

Duke Computing Cluster Manager (MCP Server)

An MCP-powered application to facilitate remote management of the Duke Computing Cluster (DCC) via SSH.

Built with:

• Remote Execution: paramiko
• Orchestration: FastMCP

🚀 Features

• SSH Management – connect to the Duke Computing Cluster using your Duke NetID + SSH key.
• File Transfer – upload and download files between your machine and the cluster.
• Remote Command Execution – run shell commands and collect outputs.
• Session Handling – disconnect and clean up connections safely.

🧩 MCP Tools

🔑 connect_to_dcc

• Connects to the DCC login node (dcc-login.oit.duke.edu by default).
• Uses Duke NetID + SSH private key authentication.
• Returns a success or failure message.

📤 upload_file

• Uploads a file from your local machine → cluster.
• Example: upload_file("./data/input.txt", "/home/<netid>/input.txt").

📥 download_file

• Downloads a file from the cluster → local machine.
• Example: download_file("/home/<netid>/results.txt", "./results/output.txt").

🖥️ execute_command

• Runs a shell command on the cluster.
• Returns: stdout, stderr, and exit_code.
• Example:

  {
    "command": "ls -l",
    "stdout": "...",
    "stderr": "",
    "exit_code": 0
  }
  
  

⚙️ Installation

📦 Using environment.yml

You can create a conda environment from the provided environment.yml file to ensure all dependencies are installed correctly.

Create the environment:

conda env create -f environment.yml

Activate the environment:

conda activate dcc-mcp

(Replace dcc-mcp with the environment name defined in environment.yml if different.)

Verify installation:

conda list

▶️ Running the MCP Server Start the MCP server:

python dcc_manager.py
This launches the FastMCP server and exposes the cluster management tools.

⚠️ Notes

Ensure your SSH key is registered with the DCC and the path matches your config.

The script currently expects a private key passphrase – change this for your setup.

Always call disconnect after work to close sessions cleanly.

🧩 LLM Integrations

The easiest LLM Agent to integrate with is Claude Desktop. Edit the claude_desktop_config.json to look something like this:

  {
   "duke-dcc-connector": {
      "command": "python",
      "args": [
        "path-to-mcp-server"
      ]
    }
  }