Skip to main content
Builds one downloadable file from either AI instructions or raw structured data. Display name “Generate File”. On by default for new agents. format: html is a static HTML file, not an interactive task-feed widget. For a live scorecard in the feed use app_tool; for a rendered view of prior step JSON use custom_interface.

​
When to use it (vs app_tool vs custom_interface vs python_code_tool)

​
Modes

AI generation — pass instructions + format. The model writes the file. Supported: html, md, csv, xlsx, pdf, docx. Do not use this when you already have structured data; it wastes an LLM call and can distort values. Direct conversion — pass content (usually a ${VAR_N} from structured_data_extraction, python_code_tool, or postgres_execute_query) + format csv, xlsx, or json. No LLM. Nested extraction metadata is stripped. For xlsx, an array of {sheetName, headers, rows} becomes a multi-sheet workbook. json is conversion-only (requires content).

​
Inputs

Required: format, filename (no extension; the tool adds it).
  • instructions — AI mode.
  • content — conversion mode; when set, instructions is ignored.
  • file_id — optional UUID of an existing file to overwrite.
  • delimiter — CSV only: , (default), ;, |, tab.
If the user uploaded files and wants a new file from their contents, extract first (structured_data_extraction / parsers), then call this tool.

​
Output

The generated file (filename, format, storage URL / file id) for download or downstream tools.

​
Limits and side effects

  • One file per call.
  • Uploads to task storage; overwriting via file_id replaces that row’s contents.
  • Repair loops may retry malformed AI output (CSV/XLSX/PDF) within configured bounds.

​
Expected errors

  • Neither instructions nor content provided.
  • format: json without content.
  • Unknown format, or conversion of data the formatter cannot shape.