Data Preprocessor MCP Server

Advanced ML-ready data preprocessing for educational and development use.

Purpose

This MCP server provides a secure, containerised interface for AI assistants to load, explore, clean, transform, and export datasets using pandas, numpy, scikit-learn, pyarrow, and openpyxl.

Features — 23 Tools

Data Loading and Inspection

• load_dataset — Load a file from your data directory (CSV, Excel, Parquet, JSON)
• get_info — Shape, column types, null counts, memory usage
• preview_data — First or last N rows of the dataset
• get_statistics — Descriptive statistics (count, mean, std, min, max, etc.)
• get_missing_summary — Per-column missing value count and percentage

Data Cleaning

• drop_columns — Remove one or more columns
• select_columns — Keep only specified columns
• drop_duplicates — Remove duplicate rows, optionally on a subset of columns
• drop_rows_with_missing — Drop rows that contain null values
• fill_missing — Fill nulls using mean, median, mode, constant, ffill, or bfill
• rename_column — Rename a single column
• filter_rows — Filter rows using:
  • ==
  • !=
  • >
  • <
  • >=
  • <=
  • contains
  • startswith
  • endswith
  • isnull
  • notnull
• detect_outliers — Detect, and optionally remove, outliers via IQR or Z-score

Type Conversion and Encoding

• convert_column_type — Cast a column to int, float, str, bool, or datetime
• encode_categorical — Label encoding or one-hot (dummy) encoding

Normalisation and Transforms

• normalize_column — Min-max (0–1) or Z-score (mean = 0, std = 1) scaling
• apply_log_transform — Apply log or log1p to a numeric column

Feature Engineering

• feature_engineering_date — Extract year, month, day, hour, minute, dayofweek, quarter, weekofyear, and dayofyear from a datetime column
• bin_column — Bin a numeric column using equal-width or quantile strategy

Sorting and Export

• sort_data — Sort dataset by a column in ascending or descending order
• export_dataset — Export to CSV, Excel, Parquet, or JSON
• reset_dataset — Reset dataset to its originally loaded state

Prerequisites

• Docker Desktop, with Docker MCP Toolkit enabled
• Docker MCP CLI plugin, with the docker mcp command available
• A local data directory to mount into the container

Installation

Step 1 — Create Your Data Directory

mkdir -p ~/mcp-data
# Copy datasets you want to work with into this directory
cp mydata.csv ~/mcp-data/

Step 2 — Build the Docker Image

Build the MCP server image from the project directory.

cd /path/to/data-preprocessor
docker build -t data-preprocessor-mcp-server .

Step 3 — Create the Custom Catalog

Create your custom Docker MCP catalog file.

mkdir -p ~/.docker/mcp/catalogs
nano ~/.docker/mcp/catalogs/custom.yaml

Paste the following, replacing YOUR_USERNAME with your macOS username:

version: 2
name: custom
displayName: Custom MCP Servers
registry:
  data-preprocessor:
    description: "ML-ready data preprocessing: cleaning, encoding, normalisation, feature engineering"
    title: "Data Preprocessor"
    type: server
    dateAdded: "2026-03-06T00:00:00Z"
    image: data-preprocessor-mcp-server:latest
    ref: ""
    readme: ""
    toolsUrl: ""
    source: ""
    upstream: ""
    icon: ""
    volumes:
      - source: "/Users/YOUR_USERNAME/mcp-data"
        target: /data
    tools:
      - name: load_dataset
      - name: get_info
      - name: preview_data
      - name: get_statistics
      - name: get_missing_summary
      - name: drop_columns
      - name: select_columns
      - name: drop_duplicates
      - name: drop_rows_with_missing
      - name: fill_missing
      - name: rename_column
      - name: filter_rows
      - name: encode_categorical
      - name: normalize_column
      - name: convert_column_type
      - name: feature_engineering_date
      - name: compute_correlation
      - name: detect_outliers
      - name: apply_log_transform
      - name: bin_column
      - name: sort_data
      - name: export_dataset
      - name: reset_dataset
    metadata:
      category: automation
      tags:
        - data-preprocessing
        - machine-learning
        - pandas
        - scikit-learn
        - csv
      license: MIT
      owner: local

Step 4 — Update the Registry

Open your local Docker MCP registry file:

nano ~/.docker/mcp/registry.yaml

Add the following under the existing registry: key:

data-preprocessor:
  ref: ""

Step 5 — Verify

Check that the MCP server is visible:

docker mcp server list

Usage Examples

You can send requests like:

"Load the file sales.csv from my data directory"

Uses:

load_dataset(filename="sales.csv")

"Show me the first 10 rows"

Uses:

preview_data(n_rows="10", position="head")

"Summarise missing values in the dataset"

Uses:

get_missing_summary()

"Fill missing values in the Age column using the median"

Uses:

fill_missing(column="Age", strategy="median")

"Drop the columns ID and Timestamp"

Uses:

drop_columns(columns="ID, Timestamp")

"Remove duplicate rows"

Uses:

drop_duplicates()

"Encode the Category column using label encoding"

Uses:

encode_categorical(column="Category", method="label")

"Normalise the Price column using min-max scaling"

Uses:

normalize_column(column="Price", method="minmax")

"Extract year, month and day from the OrderDate column"

Uses:

feature_engineering_date(column="OrderDate", features="year,month,day")

"Show the correlation matrix"

Uses:

compute_correlation()

"Detect outliers in the Revenue column using IQR"

Uses:

detect_outliers(column="Revenue", method="iqr")

"Remove outliers from the Revenue column"

Uses:

detect_outliers(column="Revenue", method="iqr", remove="true")

"Apply log1p transform to the Salary column"

Uses:

apply_log_transform(column="Salary", method="log1p")

"Bin the Age column into 4 equal-width groups"

Uses:

bin_column(column="Age", bins="4", strategy="equal_width")

"Export the cleaned dataset as cleaned_sales.csv"

Uses:

export_dataset(filename="cleaned_sales.csv", file_type="csv")

"Reset to the original dataset"

Uses:

reset_dataset()

Architecture

AI Client → MCP Gateway → Data Preprocessor Container → /data volume
                                     ↕
                            ~/mcp-data (your local files)

All data stays local. No internet access is required. No API keys are needed.

Development

Local Testing (Without Docker)

export DATA_DIR="$(pwd)/test-data"
export MAX_PREVIEW_ROWS=20
mkdir -p test-data
python data_preprocessor_server.py

Rebuild After Changes

docker build -t data-preprocessor-mcp-server .

Adding New Tools

1. Add a function with the @mcp.tool() decorator
2. Use a single-line docstring only, because multi-line docstrings can cause gateway panic
3. Default all parameters to "", never None
4. Return a formatted string
5. Add the tool name to your custom.yaml catalog entry
6. Rebuild the Docker image

License

MIT License