Convert Excel to CSV

What the CSV actually contains

"As shown" does not ask Excel what a cell looks like — it re-renders the number format here, with our own implementation of Excel's format codes, because the usual library gets hundreds of real-world formats wrong. The exception is a file that already carries its display text: an OpenDocument sheet, or one of the HTML tables that web apps name .xls, where the stored text is used as it stands. Dates never leave as serial numbers in either mode. The workbook's own epoch flag is read, so a file written by the old Mac Excel is shifted the 1,462 days it needs, and the day Excel believes existed — 29 February 1900 — is reproduced exactly where it belongs rather than sliding everything after it by one.

Three defaults are worth checking before you hand the file to something else. The delimiter starts as whatever your browser's locale implies, so a machine set to a comma decimal mark offers you semicolons. Lines end with a carriage return and a line feed unless you choose otherwise, and there is always a line break after the last row. Quoting is minimal — a field is quoted only when it contains the delimiter, a quotation mark, a line break, or leading or trailing spaces — with one special case: a first cell reading "sep=" is quoted so that Excel does not mistake your data for a delimiter instruction.

Merged cells resolve one of two ways and the default is the quieter one: only the top-left cell of a merge carries the value, and the rest come out empty. Booleans are written as TRUE and FALSE whichever mode you pick, and a cell holding an error comes out as its error text. Nothing is done to defuse a cell that begins with an equals sign. A CSV written here reproduces that text exactly as the workbook held it, so if the sheet came from somewhere you do not trust, the formula is still a formula to whatever opens the file next — which is a property of CSV rather than of this tool, and a reason to open an unknown export in a viewer first.

When you want something else

The workbook is read in full, with a ceiling of 200 MB on the file and a stop if its compressed contents claim to expand past two gigabytes. Past that, and for anything on a schedule, the answer is a desktop tool rather than the page: LibreOffice in headless mode converts a sheet to CSV in one command. For pulling one column out of a hundred-megabyte extract without materialising the rest, qsv and Miller read and write CSV as a stream.

And if the destination is a database or a pipeline rather than a person, ask whether CSV is the right target at all. It cannot record its own delimiter, its decimal mark or its encoding, which is why every CSV import begins with three questions nobody should have to answer. Parquet keeps the types, JSON Lines keeps the structure, and both are read directly by DuckDB, pandas and everything else that was going to load your CSV anyway.