My LLM wrapped the JSON in ```json fences — will it still work?

Yes. The parser scans for the first balanced JSON array in the text, so code fences, leading prose, and trailing commentary are all stripped automatically.

How are nested objects handled?

Nested objects are flattened with dot-notation keys, so {"user":{"name":"Ada"}} becomes a column called user.name. Arrays inside a row are serialised as JSON strings in a single cell.

What if rows have different keys?

The tool unions the keys across every object to build the full column set. Rows missing a key simply get an empty cell for it.

Does my data get uploaded?

No. Parsing, flattening, and CSV generation all run in your browser. Nothing is sent to a server.

LLM JSON Array → CSV Converter

LLM JSON array to CSV

LLMs love to return a JSON array of objects, then bury it inside markdown fences or a paragraph of explanation. This tool digs the array out of that mess, flattens nested objects, unions every key into a column set, and hands you a properly escaped CSV — ready to open in a spreadsheet or pipe into a script.

How it works

The input is scanned for the first balanced [ ... ] array so code fences and surrounding prose are ignored. The array is parsed with JSON.parse, then each object is flattened: nested objects become dot-notation keys (address.city), and nested arrays are serialised back to JSON inside a single cell. Columns come from the union of all keys across all rows. Every value is CSV-escaped — wrapped in quotes with internal quotes doubled — so commas, quotes, and newlines survive intact.

Tips and notes

Ask the model for an array of flat objects. The cleaner the JSON, the cleaner the CSV; deeply nested structures still convert but produce many dotted columns.
Inconsistent keys are fine. Missing fields become empty cells rather than breaking the export.
Numbers and booleans are preserved as text. CSV has no types, so true and 42 are written verbatim and re-typed by your spreadsheet on import.
Check the column count. A surprising number of columns usually means the model returned heterogeneous objects — useful to catch before downstream use.