Kaggle MCP

What arrismo/kaggle-mcp MCP server does

The arrismo/kaggle-mcp MCP server gives an MCP-compatible client access to selected Kaggle dataset workflows. It can search for datasets by keyword, return matching metadata, download and extract a chosen dataset, and generate a starter prompt for exploratory data analysis. The project is intended for local execution and lists Claude Desktop as an example client.

Search results can include a dataset reference, title, subtitle, download count, last-updated date, and usability rating. Dataset references use Kaggle's owner/dataset-slug format, such as kaggle/titanic. The generated EDA prompt requests basic Python analysis steps, including loading data, checking missing values, creating visualizations, and calculating summary statistics.

How it works

The server communicates over MCP stdio. An MCP client launches the local process and sends requests to its tools or prompt capability. A typical workflow starts with a keyword search, followed by selecting a reference from the returned results. The client can then request a download and ask for an EDA prompt for the same reference.

download_kaggle_dataset extracts the downloaded archive locally. If no output directory is supplied, files are placed under ./datasets/<dataset_slug>/; the datasets/ directory is created when a download is requested. A caller can instead provide a local download_path.

Setup and configuration

The project requires Python 3.10 or newer, a Kaggle account with an API token, and an MCP-compatible client. The documented setup clones the repository, creates a virtual environment, and installs dependencies with either uv sync or pip install -r requirements.txt.

Kaggle authentication can use KAGGLE_USERNAME and KAGGLE_KEY environment variables, including values loaded from a project .env file. Alternatively, place the standard kaggle.json file in Kaggle's expected configuration directory: ~/.kaggle/kaggle.json on macOS or Linux, or the corresponding .kaggle directory under the Windows user profile. The README recommends restricting permissions on macOS and Linux with chmod 600.

The documented local launch command is uv run kaggle-mcp. The server can also be started with python src/server.py, run in Docker, or launched through Smithery. Claude Desktop configuration uses the uv command with run kaggle-mcp arguments and may pass the Kaggle credentials in an env block.

Tools and capabilities

The arrismo/kaggle-mcp MCP server provides:

• search_kaggle_datasets(query): searches Kaggle and returns up to 10 matching datasets as JSON.
• download_kaggle_dataset(dataset_ref, download_path): downloads and unzips a dataset, using the default directory when no path is given.
• generate_eda_notebook(dataset_ref): produces a prompt for generating basic Python EDA code.

The server supports dataset discovery and preparation, but the supplied material describes EDA prompt generation rather than direct execution of analysis code or notebook creation.

Limitations and notes

A Kaggle account and API token are required. The server depends on the selected MCP client to launch the stdio process and interpret its tools or prompt. Downloads are local to the environment where the server runs, so an agent or client operating elsewhere will not automatically access those files. The documentation specifies a maximum of 10 results for each dataset search and does not describe additional Kaggle resources such as competitions or notebooks.