# jDataMunch MCP

An MCP server that answers questions about large CSV, Excel, Parquet and JSONL files by querying an indexed profile of the data instead of pasting the rows into a model's context window.

jDataMunch is an MCP server for coding agents and analysts that answers questions about tabular data files without loading the raw rows into the prompt. A dataset is indexed once, profiling its columns, types, cardinality, null rates and distributions, and stored locally; from then on, filters, aggregations and joins run against that index and the underlying file rather than against a copy of the data pasted into the conversation, so only results come back.

On a published benchmark using a public 255-megabyte, million-row dataset with 28 columns, describing the dataset's shape cost roughly 3,800 tokens against an estimated 111 million tokens to paste the raw file, and filtering or aggregating returned only the matching rows or grouped totals rather than the whole table. The project is free for personal use, with commercial use requiring a paid license.

Website: https://github.com/jgravelle/jdatamunch-mcp