RCT Detector Platform - Ultralytics

You run an Ultralytics MCP Server that connects Ultralytics YOLO models with N8N workflows to streamline training, inference, and dataset management. It centralizes model training, object detection, and automation in a single platform with GPU-accelerated performance and a Streamlit interface for easy operation.

How to use

Connect with the MCP client to start creating and automating computer vision workflows. Use the MCP server to train custom YOLO models, run real-time detections on images or videos, and trigger N8N workflows for data handling, reporting, or further processing. You can monitor training progress, view GPU utilization, and access visualization dashboards from a single web interface.

How to install

Prerequisites you need before installation: Docker and Docker Compose, NVIDIA Docker Runtime for GPU support, and at least 8 GB of RAM. You will also need a capable GPU with the appropriate drivers installed.

Step 1: Prepare your environment

• Install Docker: follow the standard installation steps for your operating system.
• Install Docker Compose: ensure the compose CLI is available on your system.
• Install NVIDIA drivers and the NVIDIA Container Toolkit to enable GPU support inside containers.

Step 2: Start the platform

• Windows users:
• Run the one-command setup: setup.bat.
• Linux or macOS users:
• Make the script executable: chmod +x setup.sh.
• Run the setup: ./setup.sh.
• Alternative manual setup:
• Start the services in detached mode from the orchestration file: docker-compose up --build -d.

Step 3: Verify access

• Main interface: http://localhost:8501.
• MCP server: http://localhost:8092.
• TensorBoard: http://localhost:6006.
• Jupyter: http://localhost:8888.

N8N automation with MCP

Connect N8N to the MCP server to automate workflows. The MCP server endpoint is exposed for SSE transport, enabling live updates and event-driven automation.

{
  "mcp_connection": {
    "transport": "sse",
    "endpoint": "http://localhost:8092/sse"
  }
}

Available operations via MCP tools

The MCP server provides a set of callable tools to process data, train models, and evaluate performance. Use the MCP client to invoke these tools and integrate results into your workflows.

Usage notes

• Do not refresh the page during dataset processing to avoid interrupting the upload workflow.
• Use the Live GPU stats and training progress indicators to optimize training parameters and resource usage.
• Keep the MCP endpoint secure and monitor health checks to ensure reliable automation.

Available tools

detect_objects

Real-time object detection in images using YOLO models, returning class labels, bounding boxes, and confidence scores.

train_model

Train custom YOLO models on user-provided datasets with selectable YOLO11 variants and training parameters.

evaluate_model

Assess model performance on validation data and generate metrics such as precision, recall, and mAP.

predict_batch

Process multiple images in a batch for efficient inference and consolidated results.

export_model

Convert trained models to formats like ONNX or TensorRT for deployment.

benchmark_model

Run performance benchmarks to compare inference speed and resource usage across models.

analyze_dataset

Compute dataset statistics and validation checks to ensure data quality before training.