JSON to TOON

Continue in the tool

JSON to TOON Compression

TOON, short for Token-Oriented Object Notation, is an alternative representation of JSON designed to fit into fewer tokens when that JSON becomes input for a language model.

How TOON saves tokens

The format combines YAML-style indentation for nested objects with a CSV-style tabular layout for uniform arrays: instead of repeating every field name in every item, an array of objects sharing the same shape becomes a single `name[N]{field1,field2}` header followed by one line per item, holding only the comma-separated values.
On a sample JSON with 3 users and a metadata block, this tool’s real conversion went from 422 to 189 characters (55.2% fewer) and from 148 to 73 tokens (50.7% fewer), counted with the same tokenizer library GPT models use; the accepted input limit is 300,000 characters.
The real pitfall is assuming TOON always shrinks the file: the gain depends on uniform, repetitive arrays, and for nested or heterogeneous data the line-by-line indentation can cost more space than the JSON braces would have saved; a small object with a mixed array (a number, a sub-array and an object inside the same array) came out to 57 characters in TOON against 47 in JSON, meaning it got bigger, not smaller.

Three real conversions, three different outcomes

Uniform array of objects: the ideal case

Input: {"users":[{"id":1,"name":"Ada Lovelace","role":"admin","active":true},{"id":2,"name":"Grace Hopper","role":"user","active":false}],"meta":{"source":"jkit.tools"}}
Expected output: users[2]{id,name,role,active}: 1,Ada Lovelace,admin,true 2,Grace Hopper,user,false meta: source: jkit.tools

The `users[2]{id,name,role,active}` header declares the array length and the 4 fields once; no data line repeats the field names, unlike JSON where every object carries the 4 keys again.

Array of numbers: the length bracket as a check

Input: {"numbers":[1,2,3,4,5]}
Expected output: numbers[5]: 1,2,3,4,5

The `[5]` is not decoration: it is an explicit check for the language model to confirm it read all 5 items, part of the format specification (v3.0) and one of the reasons TOON is easier to validate than loose JSON.

Heterogeneous data: when TOON grows instead of shrinking

Input: {"a":{"x":1,"y":{"z":2}},"b":[1,[2,3],{"c":4}]}
Expected output: a: x: 1 y: z: 2 b[3]: - 1 - [2]: 2,3 - c: 4

Since array `b` mixes a number, a sub-array and an object at the same level, no single tabular header is possible; each item becomes its own bulleted line, and the final result (57 characters) ends up longer than the input JSON (47 characters).

Full tool FAQ

Does the tool convert JSON to TOON and back again?

Yes. It converts JSON to TOON using the format's official implementation and also decodes TOON back into JSON. You can switch the mode manually or use automatic detection, and on every conversion the tool shows the character and token counts on both sides, with the savings as a percentage.

What is TOON?

How does TOON save tokens compared with JSON?

What are the delimiter and key folding?

Do fewer characters always mean fewer tokens?

Does converting to TOON and back preserve the data?

Does TOON automatically improve LLM performance?

When is it worth using TOON instead of JSON?

Frequently asked questions

What does the TOON acronym stand for?

Token-Oriented Object Notation, a format that represents the same data as JSON (objects, arrays, primitive values) but prioritizes spending fewer tokens when that content is sent as input to a language model.

Can TOON be converted back to JSON without losing data?

Yes, the conversion is lossless in both directions: the same `users[2]{id,name}` array that became a table turns back into exactly the original `{"id":1,"name":"Ada"}` objects, because the header already records the field names JSON needs to rebuild.

What kind of JSON saves the most tokens as TOON?

Long arrays of objects sharing the same fields, like user lists, orders, or exported spreadsheet rows; that is exactly where the tabular header replaces hundreds of repeated keys with a single field declaration.

Is there a size limit for converting JSON to TOON?

Yes, this tool accepts up to 300,000 input characters; above that, the conversion is refused before it runs, so the browser does not lock up trying to process an oversized JSON.

Next steps

How TOON saves tokens

The format combines YAML-style indentation for nested objects with a CSV-style tabular layout for uniform arrays: instead of repeating every field name in every item, an array of objects sharing the same shape becomes a single `name[N]{field1,field2}` header followed by one line per item, holding only the comma-separated values.

On a sample JSON with 3 users and a metadata block, this tool’s real conversion went from 422 to 189 characters (55.2% fewer) and from 148 to 73 tokens (50.7% fewer), counted with the same tokenizer library GPT models use; the accepted input limit is 300,000 characters.

The real pitfall is assuming TOON always shrinks the file: the gain depends on uniform, repetitive arrays, and for nested or heterogeneous data the line-by-line indentation can cost more space than the JSON braces would have saved; a small object with a mixed array (a number, a sub-array and an object inside the same array) came out to 57 characters in TOON against 47 in JSON, meaning it got bigger, not smaller.

Three real conversions, three different outcomes

Uniform array of objects: the ideal case

Input: {"users":[{"id":1,"name":"Ada Lovelace","role":"admin","active":true},{"id":2,"name":"Grace Hopper","role":"user","active":false}],"meta":{"source":"jkit.tools"}}
Expected output: users[2]{id,name,role,active}: 1,Ada Lovelace,admin,true 2,Grace Hopper,user,false meta: source: jkit.tools

The `users[2]{id,name,role,active}` header declares the array length and the 4 fields once; no data line repeats the field names, unlike JSON where every object carries the 4 keys again.

Array of numbers: the length bracket as a check

Input: {"numbers":[1,2,3,4,5]}
Expected output: numbers[5]: 1,2,3,4,5

Heterogeneous data: when TOON grows instead of shrinking

Input: {"a":{"x":1,"y":{"z":2}},"b":[1,[2,3],{"c":4}]}
Expected output: a: x: 1 y: z: 2 b[3]: - 1 - [2]: 2,3 - c: 4

Full tool FAQ

Does the tool convert JSON to TOON and back again?

What is TOON?

How does TOON save tokens compared with JSON?

What are the delimiter and key folding?

Do fewer characters always mean fewer tokens?

Does converting to TOON and back preserve the data?

Does TOON automatically improve LLM performance?

When is it worth using TOON instead of JSON?

Frequently asked questions

What does the TOON acronym stand for?

Can TOON be converted back to JSON without losing data?

What kind of JSON saves the most tokens as TOON?

Is there a size limit for converting JSON to TOON?

Yes, this tool accepts up to 300,000 input characters; above that, the conversion is refused before it runs, so the browser does not lock up trying to process an oversized JSON.