Mcp Ml Lab

io.github.rohithraju-opsPublished by GitHub account rohithraju-opsv0.1.0Updated Oct 24, 2026

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

VerifiedSTDIODesktop onlyData & AnalyticsAI & ML

Overview

AI-generated overview

Local MCP server that lets an assistant profile CSVs, define ML tasks, train and tune XGBoost/LightGBM models, and report results.

What you can do
  • Profile a CSV: shape, dtypes, nulls, summary stats and class balance
  • Register an ML task from a CSV, target column and classification or regression type
  • Train one or more models, optionally tuning hyperparameters with Optuna
  • Produce a markdown report with metrics, hyperparameters and feature importance
  • Compare several past experiments side by side
When to use it
Useful when you want an assistant to run real tabular ML experiments from natural language — profiling a dataset, comparing XGBoost and LightGBM, tuning with Optuna, or reviewing earlier runs — rather than only browsing experiment history.
Requirements
Runs as a local process over stdio, installed from PyPI (pip install mcp-ml-lab) with Python; add the mcp-ml-lab command to the client's MCP config and restart. No authentication, environment variables or headers are declared. Experiments are stored in a local SQLite file under the user's home directory.
Before you install
Training and tuning run locally and write experiment records to a local SQLite store, so runs persist across sessions; long Optuna tuning can be compute-heavy. The assistant reads CSV files you point it at, so keep sensitive datasets out of scope. No credentials or payments are involved.

Installation

In SourceWeft

  1. Open Mcp Ml Lab in the dashboard and add it to a workspace.
  2. Enable the server for the chats that should use its tools.

Desktop only via STDIO. STDIO servers start a local process, so they need the SourceWeft desktop host.

Other MCP clients

Follow the launch instructions in the repository.

README

mcp-ml-lab

Let AI agents run real ML experiments end-to-end.

[PyPI] [Python] [License]

An MCP server that gives Claude (or any MCP-aware AI agent) the ability to profile a CSV, define an ML task, tune XGBoost and LightGBM with Optuna, and produce a markdown report with feature importance — all from natural language.

Why this exists

The existing ML-related MCP servers wrap MLflow, ZenML, or Weights & Biases and expose them as read-only — agents can browse experiment history but can't actually run anything. mcp-ml-lab fills the gap: it lets agents execute the full experimentation loop from a user's natural-language request.

A user typing "train a model on titanic.csv to predict survival" should not need to know what XGBoost is, what cross-validation is, or how to write a hyperparameter search. The agent handles all of that — mcp-ml-lab is the tools layer that makes it possible.

Quick start

bash
pip install mcp-ml-lab

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

json
{  "mcpServers": {    "ml-lab": {      "command": "mcp-ml-lab"    }  }}

Restart Claude Desktop. The five tools below are now available.

Example queries

Try these in Claude Desktop with mcp-ml-lab connected:

  • "Profile this CSV and tell me if there's class imbalance"
  • "Compare XGBoost and LightGBM on titanic.csv with 60 seconds of tuning"
  • "Show me the top 10 features the winning model used"
  • "How did my last three experiments on the wine dataset compare?"

Tools

ToolWhat it does
inspect_dataProfile a CSV — shape, dtypes, nulls, summary stats, class balance
define_taskRegister an ML task (CSV + target + classification/regression)
run_experimentTrain one or more models, optionally tuning with Optuna
get_resultsMarkdown report with metrics, hyperparameters, feature importance
compare_runsSide-by-side comparison of multiple experiments

Each tool's full signature is in its docstring; they self-document to the LLM.

How it works

Claude Desktop  ───MCP/stdio───  mcp-ml-lab server                                  │                                  ├── data.py        CSV loading, schema inference, preprocessor                                  ├── trainers/      Pluggable XGBoost + LightGBM adapters                                  ├── search.py      Stratified CV + Optuna TPE tuning                                  ├── metrics.py     Accuracy, F1, AUC, log loss                                  ├── storage.py     SQLite via SQLAlchemy 2.0                                  └── reporting.py   Markdown report generation

All experiments and trials are persisted to ~/.mcp-ml-lab/store.db so an agent can refer back to runs across sessions.

Full design notes in ARCHITECTURE.md.

Roadmap

v0.1.0 ships classification with XGBoost and LightGBM. Planned for v0.2.0+:

  • Regression tasks
  • Time series forecasting (sktime / darts integration)
  • Deep learning baselines (pytorch-tabular)
  • Optuna multi-objective search (accuracy × latency × model size)
  • Persisted model artifacts with Docker reproducibility
  • Permutation feature importance (bias-free alternative to gain importance)
  • Notebook export — emit a Jupyter notebook that reproduces the winning run

Issues and PRs welcome.

Development

bash
git clone https://github.com/rohithraju-ops/mcp-ml-lab.gitcd mcp-ml-labpython -m venv .venv && source .venv/bin/activatepip install -e ".[dev]"pytest -v

Local debugging is easiest with the MCP Inspector:

bash
npx @modelcontextprotocol/inspector mcp-ml-lab

License

MIT.

Source: README.md at commit 486c204

Tools

0
SourceWeft has not started this package in its sandbox yet, so its tools are not known.

Version history

1
  1. v0.1.0LatestSep 16, 2026

Score

50% of the applicable checks met

How complete this listing is, computed from the registry entry, the endpoint and public sources. Not a rating of the server and not a statement that it is safe or working today: reachability is a dated check, and no installs or reviews go into the score.

  • Reached in the last 30 days25 pts

    Keep the endpoint answering the MCP handshake; SourceWeft probes it every 30 days.

  • Auth declared10 pts

    Declare the credential headers in the registry entry, or keep the endpoint answering the probe so its auth can be classified.

  • Tools listed15 pts

    Answer tools/list without a credential, or ship a package that starts in a sandbox without one.

  • DescriptionRequired15 pts
  • Logo10 pts
  • Documentation15 pts
  • License10 pts