Skip to content

DOCKTOR.md

Latest commit

 

History

History
92 lines (61 loc) · 2.75 KB

DOCKTOR.md

File metadata and controls

92 lines (61 loc) · 2.75 KB

MCPX Aggregator and Docktor

This document describes the integration of MCPX and the docktor command in DHTI.

Architecture

DHTI uses MCPX (Model Context Protocol Gateway) to aggregate multiple Model Context Protocol (MCP) servers. This allows agentic tools to interact with various inference pipelines through a single gateway.

The docktor command manages these inference pipelines, which are packaged as Docker images. Each pipeline exposes its functionality as an MCP server.

Components

  1. MCPX Container: The gateway that manages and exposes tools from connected MCP servers.
  2. Inference Pipelines: Docker containers running MCP servers (e.g., for skin cancer classification).
  3. Docktor CLI: A command-line tool to add, remove, and list inference pipelines in the MCPX configuration.

Usage

Prerequisites

  • DHTI CLI installed.
  • Docker running.

Adding an Inference Pipeline

Enable Docktor in your DHTI setup:

dhti-cli compose add -m docktor
dhti-cli docker -u

To add a new pipeline, use the docktor install command:

dhti-cli docktor install my-pipeline --image my-image:latest --model-path ./path/to/models

dhti-cli docktor install sequential-thinking --image mcp/sequentialthinking --model-path ~/temp -e DEBUG=true
  • install: The operation to perform.
  • my-pipeline: A unique name for the server.
  • --image: The Docker image containing the MCP server.
  • --model-path: (Optional) Local path to a directory containing models. This will be mounted to /model inside the container.

MCPX interface is at

Connection Details

{
	"servers": {
        "mcpx": {
                "transport": "http",
                "url": "http://0.0.0.0:9000/mcp",
        }
    }
}

Removing an Inference Pipeline

To remove a pipeline:

dhti-cli docktor remove my-pipeline

Listing Pipelines

To list all installed pipelines:

dhti-cli docktor list

Creating an Inference Pipeline Docker Image

To create a compatible inference pipeline, follow these recommendations:

  1. Base Image: Use a lightweight Python image (e.g., python:3.9-slim).
  2. MCP Sdk: Install the mcp python package.
  3. Server Logic: Implement an MCP server that exposes your inference functions as tools.
  4. Model Loading: Load models from the /model directory.
  5. Data Handling: Data can be obtained from the FHIR server or added as filename statrting with patientId in the /model directory. The input to all mcp tools should include the patientId whenever possible.
  6. Entrypoint: Set the container entrypoint to run your server script.

Examples

  • See /docktor directory for complete examples.