toon-reference

TOON Reference

TOON (Token-Oriented Object Notation) is a compact, lossless re-encoding of the JSON data model designed for LLM input. It combines YAML-style indentation for nesting with CSV-style tables for uniform arrays. Spec: https://github.com/toon-format/spec. Reference implementation: https://github.com/toon-format/toon.

Why convert to TOON

• ~40% fewer tokens than pretty JSON on mixed-structure benchmarks; comparable accuracy or better.
• Lossless for the JSON data model — round-trips back to the same JSON.
• LLM-friendly guardrails — [N] array lengths and {fields} headers give the model an explicit schema.

When TOON helps

• Uniform arrays of objects (table-shaped data) — biggest savings.
• Mixed structures with some uniform arrays.
• Anything you'd otherwise drop into a prompt as JSON.

When TOON does NOT help (or hurts)

• Pure flat tabular data — CSV is smaller than TOON; the structural overhead isn't worth it.
• Deeply nested, non-uniform structures — compact JSON often wins on tokens.
• Latency-critical local models — measure first; some quantized models process compact JSON faster.

If the user's data is one of those, say so before converting.

CLI invocation

The reference implementation is a Node CLI: @toon-format/toon.

Install on demand (no global pollution):

npx -y @toon-format/toon --help

Common forms:

# JSON → TOON (stdin/stdout)
cat data.json | npx -y @toon-format/toon encode > data.toon

# TOON → JSON (round-trip)
cat data.toon | npx -y @toon-format/toon decode > data.json

# File in / file out
npx -y @toon-format/toon encode --input data.json --output data.toon

If the user prefers a persistent install, suggest npm i -g @toon-format/toon and then use toon directly. Don't install globally without asking.

Output spot-check

After conversion, ALWAYS:

1. Round-trip: decode the produced .toon back to JSON and diff (after canonicalising key order) against the original. Must be lossless.
2. Show the user a token-count delta — count tokens on the original vs the TOON output (e.g. via tiktoken or npx ttok) and report the saving in absolute and percentage terms.

If the round-trip fails, surface the diff to the user and do not silently proceed.

File extension and media type

• Extension: .toon
• Provisional media type: text/toon (UTF-8)

Storage

This plugin writes nothing persistent. If the user wants to cache conversion artefacts (rare), put them at:

${CLAUDE_USER_DATA:-${XDG_DATA_HOME:-$HOME/.local/share}/claude-plugins}/get-toony/cache/