Data Profiler Mcp

data-profiler-mcp

[CI] [PyPI] [PyPI Downloads] [Python] [Glama] [MCP Registry] [Listed in awesome-mcp-servers] [License: MIT]

An MCP server that lets an LLM understand any tabular data file: point it at a CSV, Parquet, Excel or JSON file and get schema, distributions, data-quality flags and dtype suggestions back as structured JSON.

Stop pasting df.head() and df.info() into chat. Ask your assistant "profile sales.csv" and it reads the file itself, then tells you what is in it, what is wrong with it, and how to load it more efficiently.

[data-profiler-mcp demo: one prompt returns severity-ranked data-quality flags and a memory-saving dtype plan]

Works with Claude Desktop, Claude Code, Cursor, or any MCP-compatible client.

Features

Seven focused tools, all returning clean JSON:

Supported formats: CSV, TSV, Parquet, Excel (.xlsx/.xls), JSON and JSON Lines. Large files are read up to a row cap and clearly flagged as sampled.

No dataset at hand? examples/sample.csv is a small sales export with deliberate quality issues (missing regions, a duplicate row, a constant column, whitespace padding) -- ask your assistant to "profile examples/sample.csv" and see what it flags.

Install

No install needed to try it: open the Glama server page and use Try in Browser to call the tools against a sandbox (the repo ships examples/sample.csv at /app/examples/sample.csv to profile).

Requires Python 3.10+.

Requires MCP Python SDK 2.x (mcp>=2,<3). SDK 1.x is feature-frozen upstream and no longer supported here; 0.3.0 is the last release that ran on it.

# with uv (recommended)uv tool install data-profiler-mcp
# or with pippip install data-profiler-mcp

Or run it straight from source without installing:

git clone https://github.com/haiiibin/data-profiler-mcpcd data-profiler-mcpuv run data-profiler-mcp

Configure your client

Claude Desktop

Edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\) and add:

{  "mcpServers": {    "data-profiler": {      "command": "data-profiler-mcp"    }  }}

Running from source instead of installing? Point it at the checkout:

{  "mcpServers": {    "data-profiler": {      "command": "uv",      "args": ["--directory", "/absolute/path/to/data-profiler-mcp", "run", "data-profiler-mcp"]    }  }}

Restart Claude Desktop and the tools appear under the plug icon.

Claude Code

claude mcp add data-profiler -- data-profiler-mcp

Usage

Once connected, just talk to your assistant:

• "Profile ~/data/sales_2025.csv and tell me what's in it."
• "Are there any data-quality problems in customers.parquet?"
• "Show me 20 random rows from events.jsonl."
• "Give me full stats for the revenue column, including outliers."
• "How can I shrink this DataFrame's memory usage?"
• "What changed between snapshot_jan.csv and snapshot_feb.csv?"

Example: profile_dataset

{  "file": { "name": "sample.csv", "format": "csv", "size_human": "14.2 KB" },  "shape": { "rows": 201, "columns": 13, "sampled": false },  "memory_usage_human": "78.4 KB",  "missing_summary": { "total_missing_cells": 561, "pct_missing": 21.5, "columns_with_missing": 3 },  "duplicate_rows": { "count": 1, "pct": 0.5 },  "columns": [    {      "name": "price", "dtype": "float64", "inferred_type": "float",      "non_null": 201, "null": 0, "unique": 51,      "stats": { "min": 0.0, "max": 100000.0, "mean": 521.3, "median": 24.0 }    }  ],  "quality_flags": [    "[high] empty_col: Column is entirely empty (all values missing).",    "[warning] const: Column holds a single constant value; it carries no information.",    "[warning] numeric_text: Every value parses as a number but the column is stored as text."  ]}

Example: detect_quality_issues

{  "issue_count": 8,  "severity_counts": { "high": 2, "warning": 4, "info": 2 },  "issues": [    { "column": "empty_col", "issue": "all_missing", "severity": "high",      "detail": "Column is entirely empty (all values missing)." },    { "column": "numeric_text", "issue": "numeric_stored_as_text", "severity": "warning",      "detail": "Every value parses as a number but the column is stored as text." }  ]}

How it works

The server is built on the MCP Python SDK's MCPServer and reads files with pandas (plus pyarrow for Parquet and openpyxl for Excel). Every tool returns a plain, JSON-serializable dict, with NumPy scalars, NaN/inf and timestamps normalized so the output is safe to hand straight back to a model. Nothing is written to disk and no data leaves your machine.

Development

uv venvuv pip install -e ".[dev]"uv run pytest

License

MIT. See LICENSE.