CSV to JSON Converter

CSV is a table: one line per record, values separated by a delimiter, and usually a first row naming the columns. JSON is a structure: named keys, real types, and nesting to any depth. Converting between them means deciding what each column is called, what type each value should be, and whether any of it should nest.

1client,project,budget,active2Northwind Trading,Storefront rebuild,48000,true3 4[5  {6    "client": "Northwind Trading",7    "project": "Storefront rebuild",8    "budget": 48000,9    "active": true10  }11]

Note what happened to 48000 and true: they became a number and a boolean rather than strings. That is the part worth being careful about, and this converter is deliberately cautious about it — see below.

Everything runs in your browser. The files people convert are customer exports, order histories and internal reports, so nothing here is uploaded, stored or logged.

1. 1. Add the CSV. Paste it into the left panel, use the upload button, or drag a .csv file anywhere onto the tool. The file name appears above the panel and is reused when you download the result.
2. 2. Check the delimiter. It is detected automatically and shown in the status bar. If the file uses something unusual, pick it — or type any single character — in the options.
3. 3. Confirm the header row. The first row is treated as column names by default. Turn that off for a file that starts straight into data, and each record becomes an array instead.
4. 4. Choose how values are built. Decide whether types are detected, whether headers with dots and brackets should nest, and what an empty cell becomes.
5. 5. Read the warnings. If any row has the wrong number of columns, or a header is duplicated, the status bar says so and the Statistics tab names the line.
6. 6. Take the JSON. Copy it, or download it as a .json file named after the one you opened.

“Comma-separated” is optimistic. Exports from software configured for most of Europe use semicolons, because the comma is the decimal separator there. Database dumps often use tabs. Log pipelines use pipes. All four are built in, and any other single character can be typed in.

Detection works by consistency, not frequency. Each candidate is tried against the first twenty lines, and the one that produces the same number of columns on every line wins. Counting occurrences instead would pick the comma for the file below, which has far more commas than semicolons — and get every field wrong.

1title;note2"Hello, world, again";"one, two, three"3"a, b, c";"d, e, f"

Quoted fields follow RFC 4180. A field wrapped in double quotes may contain the delimiter, line breaks, or a doubled quote "" meaning one literal quote character. The surrounding quotes are not part of the value.

1id,company,note21,"Northwind Trading, Ltd","Ran overnight, no downtime"32,"Lumen Health","Asked for a ""dark mode"" option"

Line endings are equally varied and equally handled: \r\n from Windows, \n from everywhere else, and a bare \r from very old Mac exports all end a row.

A CSV is flat, but the JSON you want usually is not. Turn on nested JSON and the column names themselves describe the shape: a dot opens an object, and a bracketed number opens an array.

1id,customer.name,customer.email,tags[0],tags[1]21001,Mei Tanaka,mei@example.com,priority,renewal3 4[5  {6    "id": 1001,7    "customer": { "name": "Mei Tanaka", "email": "mei@example.com" },8    "tags": ["priority", "renewal"]9  }10]

The two can be combined to any depth — address.lines[0], a.b.c.d — and the containers are created as needed. If two headers ask for the same path in incompatible ways, such as a and a.b in one file, the first column to claim it keeps it and the other is skipped. Overwriting would mean silently discarding one of the two columns.

Once the shape gets deep enough to be hard to read in one block, paste the output into the JSON viewer and walk the tree, or the JSON formatter to re-indent it.

A CSV has no types. Everything in the file is text, and deciding that 42 is a number is a guess — usually a helpful one, occasionally a destructive one. The rule here is deliberately narrow:

A value becomes a number only if converting it back produces exactly the original text. 42 passes. 007 does not, because it would come back as 7 and the leading zero is gone for good. That single rule covers the cases that cause real damage:

• Product and zip codes. 007 and 01234 stay strings.
• Phone numbers. +44 20 7946 0958 is not a quantity and keeps its formatting.
• Identifiers longer than 15 digits. Beyond that, JSON numbers lose precision, so the string is the honest answer.
• Versions and dates. 1.2.3 and 2026-08-03 are text.
• Currency. $9.99 keeps its symbol.

true, false and null are recognised regardless of case. If you would rather have none of this, turn type detection off and every value stays exactly the string it was in the file.

• Check the delimiter before you trust the output. The status bar names the one in use. If a record came out as a single long field, the delimiter is wrong — not the file.
• Turn off type detection for identifiers. If a column holds account numbers, SKUs or anything with a leading zero that the narrow rule does not already protect, keeping everything as strings is safer than checking afterwards.
• Read the warnings rather than skimming past them. A ragged row almost always means a stray delimiter inside an unquoted field, and that shifts every value after it into the wrong column.
• Give every column a distinct name. Two columns called id produce id and id_2, which works but is not what anyone downstream expects.
• Decide what an empty cell means before converting. An empty string, a null and an absent key behave differently in most systems, and it is much cheaper to choose now than to patch the JSON later.
• Keep the CSV as the source of truth. For data that people edit by hand — translations, pricing, fixtures — edit the spreadsheet and convert again. Editing generated JSON by hand is how the two drift apart, and going back the other way with the JSON to CSV converter flattens nesting that a spreadsheet cannot represent.