jDataMunch MCP lets agents query million-row CSVs without pasting them into context

For a coding agent, a 255 MB CSV is not a file to read — it is a token bill. jDataMunch MCP is a local-first MCP server aimed at coding agents and analysts that answers questions about CSV, Excel, Parquet and JSONL files without pasting the rows into the context window. You index a dataset once, then retrieve column profiles, filtered rows, server-side aggregations and cross-dataset joins, with only the results returned to the model. The mechanism is straightforward and well scoped. The server profiles a file once — columns, types, cardinality, null rates and distributions — and stores that locally. Subsequent queries run against the data rather than against a copy in the prompt. Orientation questions are answered from the profile, row-level questions return matching rows, and the raw file stays out of context. That separation is the core engineering idea, and it is also where the limits sit: the value scales with file size, and the README is explicit that a 200-row spreadsheet offers little meaningful saving. Evidence in the repo Readiness signals are concrete rather than promotional. The package is distributed via PyPI, runs as a stdio MCP server with no CLI subcommands, and requires Python 3.10+ and any MCP-compatible client. The documented Claude Code setup is a single registration command using uvx jdatamunch-mcp, with persistent alternatives via uv tool install or pipx, and a warning that plain pip install is refused on PEP 668 systems. Excel and Parquet support come through optional extras, which is a sensible way to keep the base install lean. The