vaibhavparihar/chatgpt-mcp-data-quality-assistant

chatgpt-mcp-data-quality-assistant

This project is a local Model Context Protocol (MCP) server that lets ChatGPT run data quality checks and exploratory data analysis (EDA) on my local datasets via natural language.

I expose the server over HTTP (using ngrok) and register it as a custom MCP connection in ChatGPT. Once connected, I can ask ChatGPT things like:

• “Use the Local Data Quality & EDA connector and list the available datasets.”
• “Profile the titanic.csv dataset and summarise the main findings.”
• “Run data quality checks on titanic and give me the issues and a quality score.”
• “Open the profiling report for titanic and summarise it.”

ChatGPT then calls my local MCP tools, which are implemented in Python on top of pandas and profiling libraries, and returns the results back into the chat.

---

Features

• MCP server implemented in Python, compatible with ChatGPT’s MCP connections.
• Dataset discovery: lists available datasets from the local data/ folder.
• Automatic EDA:
  • Generates a profiling report for titanic.csv with summary statistics, missing values, and correlations.
  • Saves the report into reports/ as Markdown.
• Data quality checks:
  • Basic checks on missing values, numeric ranges, and simple consistency rules.
  • Returns a data quality score and a list of issues.
• Local-first: all data stays on my machine; ChatGPT only sees the results that the MCP server returns.

---

Tech Stack

• Python
• pandas for data manipulation
• Profiling / EDA libraries for automated dataset reports
• FastAPI / Starlette + Server-Sent Events (SSE) to implement the MCP transport
• ngrok to expose the local MCP server to ChatGPT
• Model Context Protocol (MCP) for the tool definitions and schema

---

How it works

1. Core logic (core.py)

   • Loads datasets from the data/ directory.
   • For titanic.csv, it generates a profiling report with:
     • Row/column counts (e.g. 887 rows, 8 columns for this Titanic file).
     • Summary statistics for numeric columns (Age, Fare, etc.).
     • Correlations between variables such as Survived, Pclass, Fare, etc. :contentReference[oaicite:1]{index=1}
   • Implements functions like:
     • list_datasets()
     • profile_dataset(name)
     • run_quality_checks(name) (returns issues + a quality score)
     • get_report(name) (loads the Markdown report for ChatGPT to summarise)

2. MCP server (server.py)

   • Wraps the core functions as MCP tools.
   • Exposes them over HTTP using FastAPI + SSE (text/event-stream), so that ChatGPT can connect as an MCP client.
   • Uses JSON-RPC 2.0 under the hood to conform to MCP expectations.

3. Exposing it to ChatGPT

   • Start the MCP server locally:

     python server.py
     

   • Expose it via ngrok:

     ngrok http 8000
     

   • Use the https://<your-subdomain>.ngrok.app/mcp URL in the ChatGPT MCP configuration, with Accept: text/event-stream.
   • Create a new MCP connection in ChatGPT and point it to this URL.

4. Using it from ChatGPT

   • Example prompts once the connector is enabled:
     • “Use the Local Data Quality & EDA connector and list the available datasets.”
     • “Profile the titanic.csv dataset and summarise the main findings.”
     • “Run data quality checks on titanic and give me the issues and the quality score.”
     • “Open the profiling report for titanic and summarise it.”

---

Example: Titanic dataset

As a demo, I use a Titanic passenger dataset stored at data/titanic.csv.

The profiling report includes:

• 887 rows and 8 columns with fields like Survived, Pclass, Age, Fare, and family relationships.
• Descriptive statistics (mean, std, min, max, quartiles).
• Correlation matrix to see how Pclass, Age, Fare, etc. relate to survival. :contentReference[oaicite:2]{index=2}

The generated Markdown report lives in reports/titanic_profile.md and can be opened and summarised by ChatGPT via the MCP connector.

---

Setup & Run

1. Create a virtual environment and install dependencies:

   python -m venv .venv
   source .venv/bin/activate      # On macOS/Linux
   pip install -r requirements.txt