🚀 MCP DS Toolkit Server

A standalone Model Context Protocol (MCP) server that brings complete DS capabilities to AI assistants like Claude Desktop and Cursor. Build, train, and track ML models through natural language - no external services required!

🎯 What is MCP DS Toolkit Server?

MCP DS Toolkit Server enables AI assistants to perform complete data science workflows through natural language. Simply talk to Claude or Cursor about what you want to do, and it handles the entire DS pipeline - from data loading to model training and evaluation.

Key Capabilities

✅ 29 DS Tools - Complete DS workflow from data loading to model evaluation ✅ Natural Language Interface - Just describe what you want in plain English ✅ Zero Configuration - Works immediately after installation ✅ Local SQLite Tracking - No external databases or cloud services needed ✅ Cross-Platform - Works on macOS, Linux, and Windows ✅ AI Assistant Integration - Seamless with Claude Desktop and Cursor IDE

Related MCP server: Data Analytics MCP Toolkit

🌟 Why MCP DS Toolkit?

🎯 Transform Your AI Assistant into a Data Scientist

• Natural Language DS: Just describe what you want - "Load the iris dataset and train a random forest classifier"

• Complete Automation: Your AI assistant handles data preprocessing, model training, and evaluation

• Intelligent Recommendations: Get suggestions for algorithms, hyperparameters, and preprocessing steps

• Comprehensive Metrics: Detailed performance metrics, learning curves, and model comparisons

🔬 Enterprise-Ready Features

• Production-Quality Code: Generated code follows best practices and is deployment-ready

• Comprehensive Tracking: Every experiment, model, and metric is automatically tracked

• Reproducible Workflows: All operations are logged and can be reproduced

• Local-First Architecture: Your data never leaves your machine

📊 Complete Tool Suite

• Data Management: Loading, validation, profiling, cleaning, preprocessing

• Model Training: 14+ algorithms from scikit-learn with automatic hyperparameter tuning

• Experiment Tracking: SQLite-based tracking with full experiment lineage

• Performance Analysis: Learning curves, feature importance, and model comparisons

Note: Cloud storage capabilities (AWS S3, Google Cloud, Azure) are available as optional dependencies but not yet fully implemented. Current version focuses on local storage and processing.

📋 Prerequisites

Install uvx (if not already installed)

This project uses uvx for easy Python package execution. If you don't have it installed:

macOS/Linux

# Install uv first (if not installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# uvx comes bundled with uv

Windows

# Install uv first (if not installed)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

# uvx comes bundled with uv

Alternative: Using pip

If you prefer not to use uvx, you can install directly with pip:

pip install mcp-ds-toolkit-server

Then use mcp-ds-toolkit-server as the command instead of uvx mcp-ds-toolkit-server in the configuration examples below.

⚡ Quick Start

Choose your preferred AI assistant:

🤖 Claude Desktop

1. Install the Server (30 seconds)

# Using uvx (recommended)
uvx mcp-ds-toolkit-server

# Or using pip
pip install mcp-ds-toolkit-server

2. Configure Claude Desktop

Add to your Claude Desktop configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json Linux: ~/.config/Claude/claude_desktop_config.json

Basic Configuration:

{
  "mcpServers": {
    "ds-toolkit": {
      "command": "uvx",
      "args": ["mcp-ds-toolkit-server"]
    }
  }
}

With Custom Data Directory:

{
  "mcpServers": {
    "ds-toolkit": {
      "command": "uvx",
      "args": [
        "mcp-ds-toolkit-server",
        "--mcp-dir",
        "/path/to/your/mcp-data"
      ]
    }
  }
}

Note: If you get uvx ENOENT errors, use the full path to uvx instead. Find it with which uvx and replace "command": "uvx" with "command": "/full/path/to/uvx". See troubleshooting section for details.

3. Restart Claude Desktop and Test

You: Load the iris dataset and train a random forest classifier
Claude: I'll help you load the iris dataset and train a random forest classifier...

📝 Cursor IDE

1. Install the Server

# Using uvx (recommended)
uvx mcp-ds-toolkit-server

# Or using pip
pip install mcp-ds-toolkit-server

2. Configure Cursor

Create or edit the MCP configuration file:

Project-specific: .cursor/mcp.json (in your project root) Global: ~/.cursor/mcp.json (in your home directory)

Basic Configuration:

{
  "mcpServers": {
    "ds-toolkit": {
      "command": "uvx",
      "args": ["mcp-ds-toolkit-server"]
    }
  }
}

With Custom Data Directory:

{
  "mcpServers": {
    "ds-toolkit": {
      "command": "uvx",
      "args": [
        "mcp-ds-toolkit-server",
        "--mcp-dir",
        "/path/to/your/mcp-data"
      ]
    }
  }
}

Note: If you get uvx ENOENT errors, use the full path to uvx instead. Find it with which uvx and replace "command": "uvx" with "command": "/full/path/to/uvx". See troubleshooting section for details.

3. Restart Cursor and Test

Open Cursor's AI chat and try:

You: Profile my CSV dataset and show me the correlations
Cursor: I'll analyze your CSV dataset and generate a comprehensive profile...

🐳 Alternative Installation Methods

Development Installation (for contributors)

git clone https://github.com/Yasserelhaddar/MCP-DS-Toolkit-Server
cd mcp-ds-toolkit-server
uv sync

# Configure with local path
{
  "mcpServers": {
    "ds-toolkit": {
      "command": "uv",
      "args": ["--directory", "/path/to/mcp-ds-toolkit-server", "run", "mcp-ds-toolkit-server"]
    }
  }
}

🛠️ Complete Tool Reference

📊 Data Management Tools (13 tools)

🤖 Model Training Tools (6 tools)

📈 Experiment Tracking Tools (10 tools)

💬 Example Prompts

🎯 Quick Start Examples

Basic Operations:
  "Load the iris dataset from sklearn"
  "Show me what datasets are currently loaded"
  "Profile my dataset and show key statistics"
  "Train a random forest classifier on the iris data"
  "Evaluate my model and show the confusion matrix"

Data Processing:
  "Load data.csv and check for missing values"
  "Clean the dataset by removing outliers using IQR method"
  "Preprocess the data with standard scaling and one-hot encoding"
  "Split my data into 70% train, 15% validation, 15% test"
  "Handle class imbalance in my dataset"

Model Training:
  "Train multiple models and compare their performance"
  "Perform 5-fold cross-validation on my model"
  "Tune hyperparameters for the random forest using grid search"
  "Show me the feature importance for the trained model"
  "Generate learning curves to check for overfitting"

Experiment Tracking:
  "Create an experiment called 'customer_churn_prediction'"
  "Start tracking this training run"
  "Log the model parameters and performance metrics"
  "Compare the last 5 training runs"
  "Show me all experiments from the past week"

🔬 Complete Workflow Example

Here's a real conversation showing an end-to-end ML pipeline:

You: Load the wine dataset from sklearn and give me a comprehensive profile

Claude: I'll load the wine dataset and provide you with a comprehensive profile...
[Loads dataset, generates statistics, correlations, and data quality report]

You: Clean the data and split it for training with 80/20 ratio

Claude: I'll clean the dataset and create train/test splits...
[Handles missing values, removes outliers, splits data]

You: Train and compare Random Forest, SVM, and Gradient Boosting classifiers

Claude: I'll train these three models and compare their performance...
[Trains models, evaluates each, provides comparison table]

You: The Random Forest looks best. Can you tune its hyperparameters?

Claude: I'll optimize the Random Forest hyperparameters using grid search...
[Performs grid search, shows best parameters and improved metrics]

You: Great! Create an experiment called "wine_classification" and log everything

Claude: I'll create the experiment and log all our work...
[Creates experiment, logs parameters, metrics, and artifacts]

You: Show me the learning curves and feature importance

Claude: I'll analyze the model's learning behavior and feature importance...
[Provides learning curve analysis and feature importance rankings]

🚀 Supported ML Algorithms

Classification Algorithms

• Tree-Based: RandomForest, GradientBoosting, ExtraTrees, DecisionTree

• Linear Models: LogisticRegression, RidgeClassifier, SGDClassifier

• Support Vector: SVC (linear, rbf, poly kernels)

• Neighbors: KNeighborsClassifier

• Naive Bayes: GaussianNB, MultinomialNB, BernoulliNB

Regression Algorithms

• Tree-Based: RandomForestRegressor, GradientBoostingRegressor, ExtraTreesRegressor

• Linear Models: LinearRegression, Ridge, Lasso, ElasticNet

• Support Vector: SVR (linear, rbf, poly kernels)

• Neighbors: KNeighborsRegressor

🏗️ Architecture

How It Works

graph LR
    A[AI Assistant<br/>Claude/Cursor] -->|Natural Language| B[MCP Protocol]
    B --> C[MCP DS Toolkit Server]
    C --> D[Data Tools]
    C --> E[Training Tools]
    C --> F[Tracking Tools]
    D --> G[Local Storage<br/>~/.mcp-ds-toolkit]
    E --> G
    F --> G

Storage Structure

~/.mcp-ds-toolkit/
├── experiments.db          # SQLite experiment tracking
├── artifacts/              # Plots, reports, outputs
│   └── {experiment_id}/
│       └── {run_id}/
├── models/                 # Saved ML models
├── datasets/               # Cached datasets
└── cache/                  # Temporary files

Technology Stack

• Core: Python 3.12+, MCP Protocol, SQLite

• ML Framework: scikit-learn, pandas, numpy

• Data Processing: pandas, numpy, scipy

• No External Dependencies: Everything runs locally

🔧 Troubleshooting

Common Issues and Solutions

Server Not Starting

# Check Python version (requires 3.12+)
python --version

# Reinstall with verbose output
pip install --verbose mcp-ds-toolkit-server

# Check if the command is available
which mcp-ds-toolkit-server

uvx Command Not Found (ENOENT Error)

If you see errors like spawn uvx ENOENT in Claude Desktop logs, this means uvx is not in the system PATH that Claude Desktop can access.

Solution: Use the full path to uvx in your configuration:

1. Find your uvx path:

   which uvx
   # Example output: /Users/username/.pyenv/shims/uvx

2. Update your configuration with the full path:

   {
     "mcpServers": {
       "ds-toolkit": {
         "command": "/Users/username/.pyenv/shims/uvx",
         "args": ["mcp-ds-toolkit-server"]
       }
     }
   }

Why this happens: Claude Desktop runs with a limited PATH environment that may not include directories where uvx is installed (like ~/.pyenv/shims for pyenv users, ~/.local/bin, or other Python tool directories).

Claude/Cursor Not Finding Tools

1. Check configuration file location:

   • Claude: ~/Library/Application Support/Claude/claude_desktop_config.json

   • Cursor: .cursor/mcp.json or ~/.cursor/mcp.json

2. Verify JSON syntax:

   {
     "mcpServers": {
       "ds-toolkit": {
         "command": "uvx",
         "args": ["mcp-ds-toolkit-server"]
       }
     }
   }

3. Restart the application after configuration changes

Permission Errors

# Fix permissions for local storage
chmod -R 755 ~/.mcp-ds-toolkit

# If using pip install
pip install --user mcp-ds-toolkit-server

Memory Issues with Large Datasets

• Use sample_dataset to work with smaller subsets

• Clear unused datasets with remove_dataset or clear_all_data

• Increase Python memory limit if needed

Getting Help

• Issues: Report bugs on GitHub Issues

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Development Setup

git clone https://github.com/Yasserelhaddar/MCP-DS-Toolkit-Server
cd mcp-ds-toolkit-server
uv sync
uv run pytest

📄 License

MIT License - see LICENSE file for details.

🙏 Acknowledgments

Built on these excellent projects:

• Model Context Protocol by Anthropic

• scikit-learn for ML algorithms

• SQLite for local tracking

Transform your AI assistant into a complete Data Science toolkit!