Overview

To let an AI agent reason about a data file, you normally paste rows into the conversation. That is expensive, truncates on anything large, and still leaves the model guessing at column types and null rates.

This server answers the question directly. One tool call returns a structured summary that is orders of magnitude smaller than the data and says more about it:

The 645 MB file cannot go into a context window at any price. It is fully characterised here in under two seconds.

Demo

python demo.py generates a deliberately messy dataset, profiles it, and reports what came back:

The same profiler seen from an MCP client — one question, one profile_dataset call, and the file is characterised without a single row entering the conversation:

Three real problems surfaced before any analysis began: a column that never varies, one that is entirely empty, and a date column that sorts as text — so "2024-10-01" < "2024-9-01" — silently corrupting any time-based result.

Features

• Six data-quality flags — constant, all-null, probable ID, mixed types, and numbers or dates stored as text.
• Full column statistics — dtype, null count and percentage, distinct count, sample values, quartiles for numerics, ranges for dates, frequent values for categories.
• Bounded output — the response stays small no matter how wide the input, and always reports what it truncated.
• Honest sampling — large files are sampled, but never silently; the true row count is always included.
• Five formats, eleven extensions — .csv .tsv .txt .parquet .pq .json .jsonl .ndjson .xlsx .xlsm .xls, plus .gz variants of the text formats.
• Path confinement — optional --root restricts profiling to a single directory.
• Zero configuration — no database, no index, no warm-up. Point it at a file.

Architecture

flowchart LR
    A["MCP client<br/>Claude Code, Claude Desktop"]
    B["server.py<br/>MCP adapter"]
    C["profiler.py<br/>pure pandas, no MCP"]
    D[("Local files<br/>CSV, Parquet<br/>JSON, Excel")]

    A -->|"profile_dataset(path)"| B
    B -->|"validate, confine to --root"| C
    C -->|"sampled read"| D
    D -->|"DataFrame"| C
    C -->|"bounded JSON profile"| B
    B -->|"tool result"| A

All profiling logic lives in profiler.py, which imports nothing from MCP. It is unit-testable without a protocol harness and usable as an ordinary Python library. server.py is only the adapter.

Installation

Requires Python 3.10+.

pip install mcp-data-profiler

pip install git+https://github.com/Ridadata/mcp-data-profiler.git

Claude Code

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

Claude Desktop and other MCP clients

Add to your client's MCP configuration:

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

To confine the server to one directory, add "args": ["--root", "/path/to/your/data"].

Usage

Once registered, ask in plain language:

• "Profile data/orders.csv"
• "Which columns have missing values?"
• "Is this dataset clean enough to model?"

Tool reference

profile_dataset(path, sample_rows=50000, max_columns=100, top_k=5, sheet=None)

Quality flags

As a Python library

from mcp_data_profiler import profile_dataset

profile = profile_dataset("data/orders.csv", sample_rows=None)
print(profile["shape"])          # {'rows_profiled': 5000, 'total_rows': 5000, 'columns': 6}
print(profile["duplicate_rows"]) # 0

Example output

Verbatim output for the sample dataset produced by python demo.py, with three of the six columns shown:

{
  "file": { "name": "orders.csv", "format": "csv", "size_bytes": 247263 },
  "shape": { "rows_profiled": 5000, "total_rows": 5000, "columns": 6 },
  "sampled": false,
  "columns": [
    {
      "name": "order_id",
      "dtype": "str",
      "null_count": 0,
      "null_pct": 0.0,
      "unique_count": 5000,
      "sample_values": ["ORD-000000", "ORD-000001", "ORD-000002"],
      "flags": ["high_cardinality_possible_id"]
    },
    {
      "name": "amount_eur",
      "dtype": "float64",
      "null_count": 0,
      "null_pct": 0.0,
      "unique_count": 1368,
      "stats": {
        "min": 2.65, "max": 1369.65, "mean": 684.2364, "std": 395.254451,
        "q25": 341.65, "median": 683.65, "q75": 1025.65
      },
      "sample_values": [2.65, 39.65, 76.65]
    },
    {
      "name": "currency",
      "dtype": "str",
      "null_count": 0,
      "null_pct": 0.0,
      "unique_count": 1,
      "top_values": [{ "value": "EUR", "count": 5000 }],
      "sample_values": ["EUR", "EUR", "EUR"],
      "flags": ["constant"]
    }
  ],
  "duplicate_rows": 0
}

Note that order_id carries no top_values: for a near-unique column every count would be 1, so the list is omitted rather than padding the response with noise.

When a file is sampled, the profile also carries "sampled": true, the true total_rows, and a sampling_note saying so.

Design notes

Bounded output. The tool must cost less than the data it describes, so the response is capped regardless of input width and long strings are truncated. Near-unique columns skip the frequent-values list, since every count would be 1.

Honest sampling. Large files are profiled from a sample, but the result always carries "sampled": true alongside the true row count — a silently sampled statistic is a wrong statistic. Row counts come from Parquet metadata or a raw newline scan, never a full parse into memory.

No silent wrong answers. The same rule governs every default that could mislead. A workbook's first sheet is often a title page, so Excel profiles always name the sheet used and list the others rather than reporting an untouched sheet as a clean dataset. CSV delimiters are inferred by testing candidates for a stable column count, which handles the semicolon files common in European open data without the header-mangling that character-frequency sniffers cause. Compressed files are decompressed before either check, since inspecting gzip bytes as text yields a plausible-looking answer that is entirely wrong.

Path safety. --root confines profiling to one directory. Paths are canonicalised before the check, so .. and symlinks cannot escape it.

Limitations