# mcp-ml-lab [Health: Active]

**Category:** 🧮 Data Science Tools  
**Repository:** https://github.com/rohithraju-ops/mcp-ml-lab  
**GitHub Stars:** 0  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/mcp-ml-lab

## Description
Run end-to-end ML experiments from natural language (XGBoost, LightGBM, Optuna).

## Tools
Capabilities this server exposes over MCP:

- **inspect_data** — Profile a CSV — shape, dtypes, nulls, summary stats, class balance
- **define_task** — Register an ML task (CSV + target + classification/regression)
- **run_experiment** — Train one or more models, optionally tuning with Optuna
- **get_results** — Markdown report with metrics, hyperparameters, feature importance
- **compare_runs** — Side-by-side comparison of multiple experiments

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `uvx` (confidence: high):

```json
"mcpServers": {
  "mcp-ml-lab": {
    "command": "uvx",
    "args": ["mcp-ml-lab"]
  }
}
```

## Documentation

## What mcp-ml-lab MCP server does

The mcp-ml-lab MCP server connects an MCP-aware AI agent to a local machine-learning experimentation workflow for CSV datasets. It supports data profiling, task registration, model training, result reporting, and comparison of previous runs. The README identifies classification with XGBoost and LightGBM as the functionality shipped in version 0.1.0.

An agent can inspect a dataset before training, identify its schema and missing values, register the target column and task type, then launch one or more experiments. The workflow is intended for tabular data and is controlled through MCP tools rather than direct command-line arguments or notebook code.

## How it works

The server uses MCP over standard input and output. Its internal components load CSV files, infer schemas, prepare data, train models through XGBoost and LightGBM adapters, run cross-validation, and generate reports. Optuna can perform TPE-based hyperparameter searches when tuning is requested.

Five tools expose the workflow:

- `inspect_data` returns dataset shape, data types, null information, summary statistics, and class-balance details.
- `define_task` records a CSV, target column, and task type for later experiments.
- `run_experiment` trains one or more supported models and can invoke Optuna tuning.
- `get_results` produces a Markdown report containing metrics, hyperparameters, and feature importance.
- `compare_runs` presents multiple experiments side by side.

The listed metrics include accuracy, F1, AUC, and log loss. Experiments and tuning trials are saved in `~/.mcp-ml-lab/store.db`, allowing later requests to refer to earlier runs across sessions.

## Setup and configuration

Install the package from PyPI with:

```bash
pip install mcp-ml-lab
```

For Claude Desktop, add an `ml-lab` entry whose command is `mcp-ml-lab` to the client’s MCP server configuration, then restart the application. The setup uses the executable installed by the package and does not document required environment variables or additional configuration values.

The repository also documents an editable development setup using a Python virtual environment, `pip install -e ".[dev]"`, and `pytest -v`. MCP Inspector can be used for local debugging by launching the installed command through `npx @modelcontextprotocol/inspector mcp-ml-lab`.

## Limitations and notes

The mcp-ml-lab MCP server is documented as shipping classification support with XGBoost and LightGBM in version 0.1.0. Regression, time-series forecasting, deep-learning baselines, multi-objective Optuna searches, model artifact persistence, permutation importance, and notebook export are listed as roadmap items rather than current capabilities.

Feature importance in the current report is available as part of run results; the planned permutation-importance feature is separate from that existing output.

_Full upstream README: https://allmcps.com/mcp/mcp-ml-lab/readme_