AutoML MCP Server for Model Selection and Tuning

Overview

AutoML MCP Server for Model Selection and Tuning is an MCP (Model Context Protocol) compatible microservice that automates common end-to-end tasks for tabular machine learning projects: dataset inspection, preprocessing, model selection and hyperparameter tuning. It exposes a set of MCP tools so external clients (MCP-aware UIs or agents) can call individual capabilities — e.g., read a CSV, run feature preprocessing, compare candidate models, or run a tuning experiment — without embedding the AutoML logic into the client.

This server is useful for data scientists and developers who want to standardize and automate repetitive workflows (EDA → preprocessing → model selection → optimization) and integrate those steps into toolchains or conversational agents. The project is lightweight, built on familiar Python ML libraries, and intended to be run locally or inside a development environment that supports MCP servers.

Repository: https://github.com/emircansoftware/MCP_Server_DataScience

Features

• Data exploration and diagnostics
  • dataset summary (shape, dtypes, missing values, memory usage)
  • visual correlation matrices, outlier detection
  • CSV loading with pandas/pyarrow support
• Automated preprocessing
  • missing value handling, categorical encoding, numeric scaling
  • duplicate detection and basic data validation
  • transformations tailored for regression and classification
• Model selection and evaluation
  • built-in comparisons across multiple algorithms
  • standardized metrics (R², MAE, MSE for regression; accuracy, F1 for classification)
  • confusion matrix and side-by-side model comparison visualization
• Hyperparameter tuning
  • automated search and trial control
  • configurable scoring functions and trial budgets
• Extensible MCP endpoints
  • discrete tools (functions) exposed as MCP actions so external controllers can orchestrate workflows

Installation / Configuration

Prerequisites:

• Python 3.8+
• pip

Clone and install:

git clone https://github.com/emircansoftware/MCP_Server_DataScience.git
cd MCP_Server_DataScience
pip install -r requirements.txt

Run the MCP server (examples):

• Using the included main entrypoint:

python main.py

• Or run the server module directly (if server exposes an mcp object):

# server.py defines `mcp`
from server import mcp
mcp.run()

• Example: start with uv (if installed and preferred):

pip install uv
uv --directory $(pwd) run main.py

Configuration tips:

• Update the CSV data path in utils/read_csv_file.py (path variable) to point at your local data folder.
• If integrating with a client (e.g., Claude Desktop), add an MCP server entry in the client configuration specifying the command and args to start this server.

Project layout (high level)

Available Tools / Resources

The server exposes multiple MCP tools grouped by capability. Common callable functions include:

Data analysis:

• information_about_data(file_name)
• reading_csv(file_name)
• visualize_correlation_num(file_name)
• visualize_correlation_cat(file_name)
• visualize_correlation_final(file_name, target_column)
• visualize_outliers(file_name)
• visualize_outliers_final(file_name, target_column)

Preprocessing & feature engineering:

• preprocessing(file_name, target_column)
• before_model(file_name, target_column)
• check_duplicates(file_name)

Modeling & evaluation:

• model_selection(file_name, target_column, problem_type)
• prediction(model_name, X_new)
• external_test(file_name, model_name) — run an external evaluation

Hyperparameter tuning:

• hyperparameter_tuning(model_name, X, y, trials, scoring_metric)
• trial_management controls number of trials and search strategy

Visualization / utilities:

• feature_importance(file_name, model_name)
• visualize_results(comparison_dataframe)

(Refer to tools/all_tools.py to discover exact function signatures and parameters.)

Supported Algorithms

Examples of algorithms included in the pipeline:

Regression: Linear Regression, Ridge, Lasso, ElasticNet, Random Forest, XGBoost, SVR, KNN, CatBoost
Classification: Logistic Regression, Ridge Classifier, Random Forest, XGBoost, SVM, KNN, Decision Tree, Naive Bayes, CatBoost

Use Cases

• Quick exploratory analysis
  • Call reading_csv() and information_about_data() to get a basic EDA report for a new dataset before modeling.
• Automated model comparison
  • Use preprocessin() then model_selection() to train several algorithms and get a ranked comparison (R²/MAE for regression; accuracy/F1 for classification).
• Hyperparameter search for a single model
  • Invoke hyperparameter_tuning() with a chosen model (e.g., XGBoost or RandomForest), set the trial budget and scoring metric, and retrieve the best hyperparameters and evaluation results.
• Integration with conversational agents or dashboards
  • Because the functionality is exposed as MCP tools, an external UI or conversational agent can run end-to-end AutoML workflows via RPC-like tool calls without embedding ML code.

Next steps / tips

• Adjust preprocessing logic and feature transformations in utils/preprocessing.py to match domain specifics.
• Extend tools/all_tools.py to add custom endpoints (e.g., custom scoring, dataset versioning).
• Use the example datasets in data/ to validate workflows and tuning parameters locally before running on production data.

For full code, API signatures and examples, see the repository: https://github.com/emircansoftware/MCP_Server_DataScience