transform_columns

Rename, cast, or drop columns of a tabular file and write the result back. This is the column-level edit that edit_table deliberately does not do (that one only changes cells/rows).

This is a write tool, so it is removed under --mcp-read-only (alongside write_table, edit_table, convert, and anonymize).

When to use

• Tidy up a schema before a load: drop junk columns, rename, fix types.
• Re-type a string column to Int64 / Float64 / Date32 and convert its cells in one pass.

Input schema

Parameter	Type	Required?	Default	Description
path	string	yes	(no default)	Path to the source file
drop	string[]	no	[]	Column names to drop (applied first)
rename	object[]	no	[]	{ "from": NAME, "to": NAME } pairs (applied after drops)
cast	object[]	no	[]	{ "column": NAME, "type": ARROW_TYPE } (applied last)
output_path	string	no	overwrite path	Where to write the result; format follows its extension
unlimited	bool	no	false	Lift the 5,000,000-row file-loader cap so every row is rewritten

Operations apply in a fixed order: drop, then rename, then cast (so cast/rename refer to the post-drop column set, and cast uses the new names). type is an Arrow type name, e.g. Int64 / Float64 / Utf8 / Boolean / Date32; values that cannot be converted stay as text. Database files (SQLite / DuckDB / GeoPackage) are not valid sources or targets.

Response shape

{
  "rows_written": 1000,
  "cols_written": 4,
  "columns": [ { "name": "id", "type": "Int64" }, … ],
  "output": "/tmp/cleaned.parquet"
}

Example call

{
  "name": "transform_columns",
  "arguments": {
    "path": "/tmp/raw.csv",
    "output_path": "/tmp/cleaned.parquet",
    "drop": ["notes"],
    "rename": [ { "from": "amt", "to": "amount" } ],
    "cast": [ { "column": "amount", "type": "Float64" } ]
  }
}

See also

• edit_table: change cells / insert / delete rows in place.
• anonymize: mask / scramble column values.