👁 Image
ToonNet 1.0.4

ToonNet

ToonNet delivers a production-ready .NET implementation of TOON (Token-Oriented Object Notation), the indentation-driven format built to minimize token usage for Large Language Models while keeping telemetry and audit logs human-readable.

👁 NuGet

Purpose

TOON replaces verbose JSON punctuation with structured whitespace, explicit schema declarations, and optional length markers. The format is ideal when you need:

• Lower token counts in prompts or completions sent to LLM providers.
• Streaming-friendly logs and metrics that remain grepable.
• Deterministic payloads that can be validated even when partially read by an agent.

How the Format Works

1. Indentation signals nesting - scopes advance by spaces instead of braces, so parsers only track the current depth.
2. Tabular arrays - declare a row shape once (items[2]{sku,qty}) and then emit comma, tab, or pipe separated values.
3. Length markers - optional #3 markers tell consumers when a block will end, preventing drift in prompt streams.
4. Key folding - single-child wrappers compact into dotted paths, but can be expanded while decoding to restore fully nested objects.

Because redundant braces and keys disappear, TOON routinely shrinks uniform payloads by 30-60 percent versus well-formatted JSON.

Installation

dotnet add package ToonNet

Quick Start

Encode

using ToonNetSerializer;

var payload = new
{
 users = new[]
 {
 new { id = 1, name = "Alice", role = "admin" },
 new { id = 2, name = "Bob", role = "user" }
 }
};

string toon = ToonNet.Encode(payload);
/*
users[2]{id,name,role}:
 1,Alice,admin
 2,Bob,user
*/

Decode

using ToonNetSerializer;

const string toon = """
users[2]{id,name,role}:
 1,Alice,admin
 2,Bob,user
""";

var dynamicPayload = ToonNet.Decode(toon);
var typedPayload = ToonNet.Decode<UserList>(toon);

Log Example

level: info
ts: 2024-05-01T10:00:00Z
service: checkout
items[2]{sku,qty}:
 SKU-1,2
 SKU-2,1

The same event encoded as JSON would include repeated keys, quotes, and braces, adding roughly 60 extra characters per entry and inflating LLM token counts.

DataTable Round-Trip

using System.Data;
using ToonNetSerializer;

var table = new DataTable("Users");
table.Columns.Add("Id", typeof(int));
table.Columns.Add("Email", typeof(string));
table.Columns.Add("Active", typeof(bool));
table.Rows.Add(1, "a@example.com", true);
table.Rows.Add(2, "b@example.com", false);

var toon = ToonNet.Encode(table);
/*
tableName: "Users"
columns[3]{name,type,allowNull,unique,isPrimaryKey}:
 Id,System.Int32,true,false,false
 Email,System.String,true,false,false
 Active,System.Boolean,true,false,false
rows[2]{Id,Email,Active}:
 1,a@example.com,true
 2,b@example.com,false
*/

var roundTripped = ToonNet.Decode<DataTable>(toon);
Console.WriteLine(roundTripped!.Rows.Count); // 2
Console.WriteLine(roundTripped.Rows[0]["Email"]); // a@example.com

Configuration Surface

var encodeOptions = new ToonOptions
{
 Indent = 1,
 Delimiter = ToonDelimiter.Tab,
 UseLengthMarker = true,
 KeyFolding = KeyFoldingMode.Safe,
 FlattenDepth = int.MaxValue
};

var decodeOptions = new ToonDecodeOptions
{
 Indent = 1,
 Strict = true,
 ExpandPaths = PathExpansionMode.Safe
};

Use the same option set on both sides when you need lossless round-trips across services.

Feature Highlights

• Token-aware serialization - reduces punctuation and repeated field names to stay within LLM context windows.
• Deterministic and streamable - indentation plus optional #length markers keep agents from drifting when consuming partial data.
• Human-first logging - can be tailed like plain text yet parsed like structured data.
• Dependency-free - the package only relies on the .NET base class library.

Compatibility

• .NET 10
• .NET 9
• .NET 8
• .NET 7
• .NET 6
• .NET Standard 2.0 (covers .NET Framework 4.6.1+, .NET Core 2.0+, Mono, Unity)

License

MIT License. See .