Data Quality MCP Server

A Model Context Protocol (MCP) server that runs data-quality checks on a SQL warehouse and exposes them to an AI assistant — so anyone can ask, in plain English, "were there any duplicate PAN numbers in last night's load?" and get a real answer from live data.

Built to demonstrate a pattern I care about as a Data QA engineer: write your validation logic once, then serve it through two front doors.

The core idea: one logic layer, two front doors

The checks live in src/checks.py as plain Python functions. Nothing about them is AI-specific. That single layer is then reused by:

Same duplicate-PAN check runs in your nightly pipeline and answers a product manager's ad-hoc question. Write once; run automatically; ask conversationally.

Related MCP server: mcp-sql-api

Checks included

Classic QA — row_count, check_nulls (null/empty %), check_duplicates. Profiling — profile_column (min/max/distinct), check_freshness (row age). Discovery — list_tables, describe_table. Plus run_all_checks for the suite.

Every check returns a structured result with a PASS / FAIL / FRESH / STALE status. Table and column names are validated against the live schema before any SQL is built (prevents injection, gives clear errors).

Project structure

data-quality-mcp/
├── src/
│   ├── checks.py       # core check logic (the shared layer)
│   ├── db.py           # SQLite connection + identifier validation
│   ├── server.py       # MCP tools (front door #1: conversational)
│   └── run_checks.py   # CLI report + CI exit code (front door #2: pipeline)
├── scripts/seed_db.py  # sample warehouse with intentional issues
├── tests/test_checks.py
├── requirements.txt
└── README.md

Quick start

python -m venv venv
# Windows: venv\Scripts\activate   |   macOS/Linux: source venv/bin/activate
pip install -r requirements.txt

python scripts/seed_db.py     # build the sample DB
python -m src.run_checks      # pipeline front door — prints a report
pytest -q                     # run the tests

The sample data is seeded with deliberate problems (missing emails, a duplicate PAN, a negative order amount) so every check has something real to catch.

Front door #1 — connect to Claude Desktop

Add this to Claude Desktop's config (Settings → Developer → Edit config), using absolute paths to your venv Python and the repo:

{
  "mcpServers": {
    "data-quality": {
      "command": "/absolute/path/to/venv/bin/python",
      "args": ["-m", "src.server"],
      "cwd": "/absolute/path/to/data-quality-mcp"
    }
  }
}

Restart Claude Desktop, then ask: "List the tables, then check the customers table for duplicate PANs and tell me if any order amounts look wrong."

Front door #2 — use in CI/CD

python -m src.run_checks exits 1 if any check fails, so it drops straight into a pipeline step or an Airflow task — the run fails loudly when data quality regresses.

Tech

Python · SQLite · Model Context Protocol (MCP) · pytest

Built by Prashant (github.com/psat022) — Senior Data QA / ETL engineer, exploring the overlap between data validation and AI tooling.