Dataset Explorer

io.github.khanarmaghanrasheed-18v0.2.1更新于 Oct 8, 2026

Help your AI explore local data using summaries, Pearson correlation, eta squared, and Cramer's V.

已验证STDIO仅桌面Files & StorageData & Analytics

概览

AI 生成的概览

让助手读取本地的 CSV、Excel、JSON 或 Parquet 文件，并计算摘要、缺失值、重复行、离群值和列间关联。

功能
Dataset Explorer 在本地运行，加载你指定的数据文件，返回实际计算出的统计结果而非猜测。工具包括 get_dataset_overview、dataset_shape、dataset_statistical_summary、inspect_Column、analyze_target、duplicate_finder、analyze_missing_values、find_correlations、detect_outliers 和 screen_target_relationships。它会根据所比较列的类型选择 Pearson 相关系数、eta 平方或 Cramer's V。此外还提供 dataset://guide 资源和 explore_dataset 提示。
适用场景
当你希望助手检查本机上的数据文件时使用：查看列类型和缺失值、查找重复行、发现离群值，或了解哪些特征与目标列相关。它面向探索性分析，不用于建模或因果推断。
运行要求
本地 Python 3.10 或更高版本；可从 PyPI 安装 dataset-explorer-mcp，或用 uvx 免安装运行。需要支持本地 stdio 服务器的 MCP 客户端，例如 Claude Desktop、VS Code with Copilot 或 Cursor。数据文件必须在同一台电脑上可读，大文件需要足够内存，因为每个工具都会把数据集载入内存。
安装前请注意
服务器会读取你指定的数据文件，且不修改原文件，但不限制可读取的路径。结果可能按客户端设置发送给助手的 AI 提供商，因此除非可以接受，否则不要指向敏感文件。统计结果只反映关联，不能证明因果关系。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Dataset Explorer，将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only，通过 STDIO。 STDIO 服务会启动本地进程，因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

Dataset Explorer MCP

Dataset Explorer helps your AI assistant understand data files saved on your computer. Ask it to summarize a dataset, check missing values, find repeated rows, or spot unusual patterns. It calculates answers from your file and leaves the original unchanged.

It works with CSV, TSV, Excel, JSON, and Parquet files. Your assistant starts this small local server when needed. You don't need a web app, a hosting account, or a Gemini API key. Your assistant may send the results to its AI provider according to that client's settings.

The connection uses MCP over stdio, which simply means your assistant talks directly to the server running on your computer.

How it finds relationships

The server calculates statistics from your dataset rather than guessing them:

  • Pearson correlation checks how two numeric columns move together.
  • Eta squared compares numeric values across groups, such as scores across categories.
  • Cramer's V measures the relationship between two category columns.

It chooses the method based on the types of columns being compared. These statistics show associations; they don't prove that one feature causes another.

Install

Available on PyPI.

You need Python 3.10 or newer.

sh
python -m pip install dataset-explorer-mcp

The command to start the server is:

sh
dataset-explorer-mcp

Your MCP client normally runs this command for you. If you run it in a terminal, it waits quietly for messages from a client. That is expected.

If you use uv, you can run it without a separate installation:

sh
uvx dataset-explorer-mcp

Connect your assistant

Use an MCP client that supports local stdio servers, such as Claude Desktop, VS Code with Copilot, or Cursor. The settings file differs by client.

For Claude Desktop or Cursor, add this to your MCP configuration:

json
{  "mcpServers": {    "dataset-explorer": {      "command": "uvx",      "args": ["dataset-explorer-mcp"]    }  }}

For VS Code, use .vscode/mcp.json:

json
{  "servers": {    "dataset-explorer": {      "type": "stdio",      "command": "uvx",      "args": ["dataset-explorer-mcp"]    }  }}

If you installed with pip, use "command": "dataset-explorer-mcp" and "args": [] instead. An absolute path to the executable also works. Reload your client after changing its configuration.

Use a local file

Give your assistant the full path to your dataset. For example:

Explore C:/Users/YourName/Downloads/customers.csv. Check missing values and duplicates.

Summarize /home/yourname/data/sales.xlsx and inspect the revenue column.

In /Users/yourname/data/results.parquet, which features are associated with the target column score?

All tools take a path. A direct tool call looks like:

json
{"path": "C:/Users/YourName/Downloads/customers.csv"}

Use forward slashes in Windows paths, or double backslashes when writing JSON. The file must be available on the computer where the server runs.

Supported files and tools

Supported files: CSV, TSV, Excel (.xlsx, .xls), JSON, and Parquet. Excel reads the first worksheet. JSON must contain tabular data that Pandas can read.

ToolWhat it does
get_dataset_overviewLists columns, types, and missing-value counts
dataset_shapeCounts rows and columns
dataset_statistical_summaryCalculates numeric means and medians
inspect_ColumnSummarizes one column; also takes col_name
analyze_targetInspects a target column; also takes target_name
duplicate_finderFinds repeated rows
analyze_missing_valuesReports missing data
find_correlationsFinds related numeric columns; optional threshold defaults to 0.8
detect_outliersFinds unusual numeric values
screen_target_relationshipsCompares features with a target; also takes target

The server also offers the dataset://guide resource and an explore_dataset prompt. These results help you explore data; they don't prove causes or train a model. Large files need enough RAM because each tool loads the dataset into memory.

Troubleshooting

  • Command not found: use the full path to dataset-explorer-mcp, or install uv and use the uvx configuration above.
  • File not found: use an absolute path and check that the server can read it.
  • No tools appear: check your client's server logs and reload its MCP settings.
  • Server seems idle: it is waiting for the MCP client; connect it through your assistant rather than typing questions into the server terminal.
  • Unsupported file: save the data in one of the formats listed above.
  • Missing values in statistics: empty or constant columns may have undefined statistics. Check the overview and missing-value tools first.

Normal server output is reserved for MCP messages. Diagnostics go to stderr, which your client's server logs usually display.

Run from source

sh
git clone https://github.com/khanarmaghanrasheed-18/MCP-Dataset-Explorer.gitcd MCP-Dataset-Explorerpython -m pip install -e ".[dev]"python -m pytest -qpython mcp_server.py

Build the downloadable package with python -m build.

License

MIT. See LICENSE.

来源：README.md，提交 c5b6fbd

工具

0
工具元数据尚未被收录。

版本历史

1
  1. v0.2.1最新Oct 8, 2026