eda-mcp

An MCP server for exploratory data analysis. Point it at a dataset and let your AI assistant do the analysis — summary statistics, diagnostic plots, correlation analysis, and full markdown reports, all from a single conversation.

Built by MLMecham.

Quickstart

Run instantly with no install step:

uvx eda-mcp

Or install permanently:

pip install eda-mcp

Related MCP server: mcp-csv-analyst

Connecting to Claude Desktop

Add this to your claude_desktop_config.json:

Mac: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "eda-mcp": {
      "command": "uvx",
      "args": ["eda-mcp"]
    }
  }
}

Restart Claude Desktop. The tools will appear automatically.

Tip: Add --refresh to always pull the latest version from PyPI on startup:

"args": ["--refresh", "eda-mcp"]

Troubleshooting

Tools not appearing after install or update

uvx caches the installed version and won't update automatically. Force a refresh:

uvx --refresh eda-mcp --help

Then fully quit and reopen Claude Desktop (not just close the window).

Check server logs

If the tools still don't appear, check the MCP server logs:

• Windows: %APPDATA%\Claude\logs\mcp-server-eda-mcp.log

• Mac: ~/Library/Logs/Claude/mcp-server-eda-mcp.log

Tools

Supported File Formats

String columns are automatically coerced to better types on load (integers, floats, dates) where unambiguous.

For SQLite and DuckDB files with multiple tables, pass the table parameter to specify which one. If the database has exactly one table it is loaded automatically.

Querying with SQL

Use query_dataset for SQL-based loading, remote sources, or cross-file joins:

-- Filter before analysis
SELECT * FROM 's3://bucket/sales.parquet' WHERE year = 2024

-- Cross-file join — mix any DuckDB-readable sources
SELECT t.*, p.bst FROM 'trainers.csv' t JOIN 'pokemon.parquet' p ON t.pokemon = p.name

-- Query a local DuckDB database (pass db_path separately)
SELECT * FROM my_table

-- Hive-partitioned S3
SELECT * FROM read_parquet('s3://bucket/data/', hive_partitioning=true)

Pass the result_path from query_dataset to any other tool exactly like a regular file_path.

Association Analysis

get_correlations computes three types of associations in one call:

Toggle each type with numeric=True/False, categorical=True/False, mixed=True/False. Set plots=True to generate heatmaps and pair-level charts.

Comparing distributions

compare_distributions diffs two slices column by column:

# Compare two cut grades
compare_distributions(
    "SELECT * FROM 'diamonds.parquet' WHERE cut='Ideal'",
    "SELECT * FROM 'diamonds.parquet' WHERE cut='Fair'",
    label_a="Ideal", label_b="Fair"
)

# Compare two time periods
compare_distributions("sales_2023.parquet", "sales_2024.parquet", label_a="2023", label_b="2024")

Returns mean, median, std, outlier, and missing value deltas per column — Claude can immediately say how much each statistic changed.

Column Classifications

Every column is automatically classified before analysis:

Pass classification="categorical" to any summary or plot tool to override the auto-detected type.

Using as a Python Library

The core functions are also importable directly:

from eda_mcp import (
    load_file, load_query,
    classify_column, get_summary,
    numeric_columns, categorical_columns,
    compute_correlations, compute_cramers_v, compute_eta_squared,
    generate_markdown_report,
)

df = load_file("data/sales.parquet")
summary = get_summary(df["revenue"])
generate_markdown_report(df, "data/sales.parquet", "output/")

# Query and analyze
df = load_query("SELECT * FROM 'data.parquet' WHERE region='West'")

Example Prompts

Once connected to Claude:

Analyze this dataset: /path/to/data.csv

Join the Batting and People tables in my Lahman SQLite database and generate a full EDA report

Compare the price distribution between Ideal and Fair cut diamonds

How does revenue vary across regions and product categories?

What columns in sales.parquet have missing values?

Generate a full EDA report for customers.xlsx

Requirements

• Python 3.11+

• Dependencies are installed automatically via uvx or pip

License

MIT