JSON → TOON
Output appears here...How to use
01
Paste JSON and pick a delimiter
Drop your payload in the input pane and choose comma, tab, or pipe from the delimiter select. Tab often tokenizes cheapest; comma is easiest to read.
02
Click Convert to TOON
Uniform object arrays collapse into a header row plus one line per record. The savings panel then compares estimated tokens against the minified JSON baseline.
03
Copy it, or round-trip it
Paste the TOON into your prompt, ideally noting the format for the model. Press Swap direction any time to turn TOON back into the original JSON.
Why JSON to TOON Converter
- Cuts LLM token usage by 30-60% on structured data, lower GPT and Claude API bills.
- Lossless: TOON converts back to the exact same JSON.
- Tabular encoding writes repeated object keys once instead of every row.
- Live token-savings estimate for input vs output.
- Comma, tab, or pipe delimiters, tab often tokenizes cheapest.
- Runs 100% in your browser, your data never touches a server.
Common questions
What does the TOON output actually look like?
Indentation-based structure, no braces. Plain fields are key: value lines, primitive arrays inline like tags[3]: a,b,c, and uniform object arrays become a header such as products[4]{id,sku,name,price} followed by one delimited row per record. The declared lengths and field list also give the model something to validate against.
How much can this reduce llm prompt tokens in practice?
Expect 30 to 60 percent on flat, table-like data such as query results or product catalogs, because that's where repeated keys dominate JSON's size. Deeply nested or non-uniform data saves much less, sometimes nothing. The savings panel measures your actual payload, so you don't have to guess.
Is TOON an official standard I can rely on?
No, and it's worth being clear about that: it's a young format with a public spec maintained on GitHub, not an RFC. It's a sensible choice for prompt payloads you control end to end, and a poor choice for public API contracts. Since this tool converts both directions, adopting it isn't a one-way door.
Can the original JSON be recovered from the TOON?
Yes, exactly. Press Swap direction, paste the TOON, and the decoder rebuilds the same structure with strings, numbers, booleans, nulls, objects, and arrays intact. Round-tripping is lossless, which is what separates a real serialization format from mere prompt shorthand.
Which delimiter should I pick in the converter?
Comma is the readable default. Tab frequently wins on token count because tokenizers tend to merge a tab into one token and tab characters rarely appear inside data, so fewer cells need quoting. Pipe sits in between. The choice is declared in the array header, so decoding works the same regardless.
Do GPT and Claude actually understand TOON input?
Reliably, yes, because it reads like a blend of YAML and CSV, both heavily represented in training data. The explicit array lengths and field headers help the model keep rows straight. One line in your prompt saying the data is in TOON format removes any remaining ambiguity.
How trustworthy is the token savings estimate?
It's a heuristic modeled on GPT-style BPE tokenization, so absolute numbers will differ a bit per model. Both sides of the comparison use the same estimator, though, which makes the percentage a fair relative measure. For billing-grade numbers, run both versions through your provider's own tokenizer.
Are there cases where a toon format converter is the wrong choice?
Yes: anything consumed by a strict JSON parser, like function-calling schemas, tool outputs, or REST payloads, should stay JSON. The same goes for deeply nested config-like data, where TOON's tabular advantage disappears. Use it where it wins, which means large flat datasets pasted into prompts.