Dispatcher MCP Server

Dispatcher MCP Server

An MCP (Model Context Protocol) server that acts as a wrapper around the dpdispatcher library. It allows language models or other MCP clients to submit and manage computational jobs on local machines or HPC clusters supported by dpdispatcher.

Features

• Exposes dpdispatcher functionality via standard MCP tools.
• submit_job: Submits a new computation job.
• query_status: Checks the status of a submitted job.
• cancel_job: Attempts to cancel a running or queued job.
• fetch_result: Retrieves the paths of result files for a completed job.
• Includes MCP Resources and Prompts to guide interactive job configuration.
• Supports stdio transport for local integration (e.g., with Cline).

Setup

1. Prerequisites:

   • Python 3.x
   • dpdispatcher library installed (pip install dpdispatcher)
   • mcp library installed (pip install mcp)
   • anyio library installed (pip install anyio)

2. Clone/Place Files: Ensure fast_server.py and job_manager.py (and __init__.py) are within a directory (e.g., dispatcher_mcp_server).

3. Configure dpdispatcher: If submitting to remote HPCs or Bohrium, ensure dpdispatcher itself is correctly configured (e.g., SSH keys, Bohrium credentials).

Running the Server

Navigate to the parent directory containing dispatcher_mcp_server and run:

python dispatcher_mcp_server/fast_server.py

The server will start and listen via stdio.

Integration with MCP Clients (e.g., Cline)

Add the following configuration to your client's MCP settings (e.g., mcp_settings.json for Cline), adjusting paths as necessary:

{
  "mcpServers": {
    "dispatcher-mcp-server": {
      "command": "python",
      "args": [
        "dispatcher_mcp_server/fast_server.py"
      ],
      "cwd": "/path/to/parent/directory/containing/dispatcher_mcp_server",
      "env": {
        "PYTHONPATH": "/path/to/parent/directory/containing/dispatcher_mcp_server"
      },
      "disabled": false
    }
  }
}

Restart the client to load the server.

Usage

Interact with the server using an MCP client. You can directly call tools like submit_job by providing the necessary arguments (machine config, resources config, task details), or use the configure_job prompt to guide an LLM through an interactive configuration process. Helper resources like dpd://examples/machine/{type} are available for context.