JSON to TOON Converter

What is TOON (Token-Oriented Object Notation)?

TOON is a compact text format for the JSON data model, built for one job: putting structured data into prompts for large language models using as few tokens as possible. It was published as an open specification with a reference TypeScript implementation, and it has quickly become a common way to send tables of records — search results, database rows, product catalogues, logs — to a model.

It borrows two familiar ideas. Objects use indentation, like YAML: name: Ada on one line, nested fields indented beneath their parent. Arrays of objects that all share the same keys become a table, as in CSV: a header such as users[3]{id,name,role}: declares the length and the column names once, then each item is a single comma-separated row. Everything JSON can express — strings, numbers, booleans, null, nested objects, mixed arrays — has a TOON form, so nothing is lost in conversion.

Why TOON uses fewer tokens than JSON

Models are billed, and limited, by tokens, and JSON spends many of them on structure rather than data. In an array of 100 users, formatted JSON repeats every key 100 times and wraps every key and string in quotes, every record in braces, with commas and indentation between them all. Tokenizers count most of that punctuation as tokens.

TOON writes each key once, in the array header, and drops the quotes and braces that the header makes unnecessary. Strings are quoted only when they would otherwise be ambiguous — when they look like a number or boolean, or contain the delimiter or a colon. On uniform data that typically removes 30–60% of the tokens compared with formatted JSON, and a meaningful share compared with minified JSON. The token panel under the editors measures this on your own data with GPT's o200k tokenizer, so you don't have to take a typical figure on trust.

When TOON helps — and when JSON is the better choice

TOON is not smaller for every document. As a rule of thumb:

• Big win: arrays of objects with the same keys — API results, rows from SQL or a spreadsheet, event logs, embeddings metadata.
• Moderate win: configuration-style objects with a few levels of nesting, where indentation replaces braces and quotes.
• Little or no win: deeply nested data with few repeated structures, and arrays whose objects have different keys, which TOON writes as list items rather than a table.
• Use JSON when: the text must be parsed by something that only speaks JSON, or when the model needs to return output your code consumes directly — or ask for TOON and decode it with the TOON to JSON converter.

Comma, tab, or pipe: choosing a delimiter

The delimiter separates values in tables and inline arrays. Comma is the default and the most familiar. Tab often saves a few more tokens, because tokenizers rarely merge a tab with the text around it and real values seldom contain tabs, so fewer of them need quotes. Pipe is a readable compromise when your values are full of commas, such as addresses or prose.

Whichever you choose is written into each array header — [3\t] for tab, [3|] for pipe — so a decoder never has to guess. Switch between them above and watch the token count to find the cheapest option for your data.

Using TOON in LLM prompts

Paste TOON into a prompt where you would otherwise paste JSON, ideally inside a fenced code block labelled toon so the model sees clearly where the data starts and ends. Models read it without special instructions, because the array headers spell out the item count and the field names before the rows. If you want the model to answer in TOON too, show it the header you expect, and ask it to keep the [N] count accurate.

Then decode the reply with the TOON to JSON converter. Its strict length check catches a reply that was cut off mid-table, and the result can be checked against your expected structure with the JSON Schema validator.

TOON compared with CSV, YAML, and minified JSON

CSV is just as compact for a single flat table, but it has no types, no nesting, and no way to hold more than one table — use JSON to CSV when the destination is a spreadsheet. YAML shares TOON's indentation but repeats every key for every list item, so it saves far less on arrays; JSON to YAML is the right choice for config files people edit. Minified JSON, which the JSON formatter produces in one click, is the safe universal baseline: TOON is worth it when the token panel shows it beating that baseline, not just formatted JSON.