Introducing TOON: A Token-Efficient Format for LLM Prompts

We're excited to introduce TOON (Token-Oriented Object Notation), a compact serialization format designed specifically for sending structured JSON data to large language models. Unlike minified JSON, which saves bytes but is unreadable, or pretty-printed JSON, which is readable but token-heavy, TOON aims to give you both: fewer tokens and a format humans can still scan and edit.

🤖 The Problem with Sending JSON to LLMs

When you build AI agents, RAG pipelines, or chatbot tools, you often need to stuff lists of records — search results, database rows, API responses — directly into a prompt. JSON is the natural choice, but it repeats every field name for every object in an array, which burns tokens fast. Every token in the prompt costs money and counts against the model's context window.

{
  "users": [
    { "id": 1, "name": "Alice", "role": "admin" },
    { "id": 2, "name": "Bob", "role": "editor" },
    { "id": 3, "name": "Carol", "role": "viewer" }
  ]
}

Notice that "id", "name", and "role" are each repeated once per object. In JSON, that per-object key repetition is the main source of wasted tokens. For a large list of search results or database rows passed as LLM context, this overhead adds up quickly.

✨ Introducing TOON Format

TOON addresses this by declaring the field names for an array of uniform objects once, in a header line, and then listing each row as comma-separated values with no repeated keys. Here's the same user list in TOON:

users[3]{id,name,role}:
  1,Alice,admin
  2,Bob,editor
  3,Carol,viewer

The [3] declares the row count and {id,name,role} declares the field names once — every row afterward is pure data. TOON also supports plain nested objects with indentation instead of braces, and inline arrays of primitives with a count marker:

location:
  city: Berlin
  country: DE
tags[3]: red,green,blue

🎯 Key Features of TOON

🚀 Use Cases

TOON is particularly well-suited for:

• RAG pipelines: Stuffing retrieved documents or database rows into a prompt as context
• AI agent tool calls: Passing structured tool outputs and API responses back to the model
• Search results: Sending lists of ranked results to an LLM for summarization or reasoning
• Chatbot context: Including user data, history, or records in a conversation prompt
• Large context windows: Fitting more data into a fixed token budget

🔮 The Roadmap for TOON Tooling

TOON is an evolving, open format, and there's more tooling we'd like to see grow around it. Some general directions for this site and the broader ecosystem:

💡 Getting Started

Ready to try TOON? Here's how to get started:

1. Convert your existing JSON payloads using our online converter
2. Explore the complete documentation to learn the syntax
3. Check out the spec and reference implementation on GitHub
4. Start by converting the arrays of records you already send to an LLM, since those benefit the most from TOON's tabular format

🤝 Join the Community

TOON is an open format, and we welcome contributions from developers building AI and LLM applications. Whether you're building a RAG pipeline, an agent framework, or a chatbot, your feedback and contributions help make TOON better for everyone.