TOON MCP Server

MCP (Model Context Protocol) server for TOON (Token-Oriented Object Notation) encoding. Reduce LLM token usage by 50-70% when sending structured data.

What is TOON?

TOON is a compact data format optimized for LLM input. Instead of repeating field names for every object, it uses a header-based format:

JSON (1041 tokens):

[
  {"id": 1, "name": "Product A", "price": 99.99},
  {"id": 2, "name": "Product B", "price": 149.99}
]

TOON (389 tokens):

[id,name,price]
1,Product A,99.99
2,Product B,149.99

Result: 62% fewer tokens = 62% cost savings

Installation

As MCP Server (Claude Desktop / Claude Code)

1. Clone or download this repository

2. Install dependencies:

   cd toon-mcp
   npm install

3. Add to your MCP settings:

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "toon": {
      "command": "node",
      "args": ["/path/to/toon-mcp/src/index.js"]
    }
  }
}

Claude Code (~/.claude/settings.json):

{
  "mcpServers": {
    "toon": {
      "command": "node",
      "args": ["/path/to/toon-mcp/src/index.js"]
    }
  }
}

As Claude Code Skill

Copy skills/toon.md to your Claude Code skills directory:

cp skills/toon.md ~/.claude/skills/

Then use /toon in Claude Code.

Available Tools

toon_encode

Convert data to TOON format.

Supported formats: JSON, CSV, TSV, XML, HTML tables, YAML

Input: [{"name":"Alice","age":30},{"name":"Bob","age":25}]
Output: [name,age]
        Alice,30
        Bob,25

toon_decode

Convert TOON back to JSON.

toon_analyze

Analyze data and show potential token/cost savings.

toon_optimize_prompt

Find data sections in a prompt and convert them to TOON automatically.

Usage Examples

In Claude Desktop/Code (with MCP)

Just ask Claude to use the tools:

• "Encode this JSON to TOON: [...]"
• "Analyze how much I'd save converting this data to TOON"
• "Optimize this prompt for token efficiency"

Programmatic (Node.js)

const { ToonEncoder } = require('./src/toon-encoder.js');

// Encode
const data = [
  { id: 1, name: 'Test', price: 99.99 },
  { id: 2, name: 'Test 2', price: 149.99 },
];
const toon = ToonEncoder.encode(data);

// Get stats
const json = JSON.stringify(data);
const stats = ToonEncoder.getStats(json, toon);
console.log(stats.savings.percent); // "64.5%"

// Decode
const decoded = ToonEncoder.decode(toon);

Benchmarks

Tested with OpenAI GPT-4o-mini:

| Dataset Size | JSON Tokens | TOON Tokens | Savings | |--------------|-------------|-------------|---------| | 5 items | 383 | 192 | 49.9% | | 20 items | 1,394 | 530 | 62% | | 50 items | 3,412 | 1,204 | 64.7% | | 100 items | 6,800 | 2,400 | ~65% |

Cost Savings at Scale

| Volume | GPT-4o-mini | GPT-4o | Claude Sonnet | |--------|-------------|--------|---------------| | 1M requests | $489 saved | $8,158 saved | $9,789 saved | | 10M requests | $4,890 saved | $81,580 saved | $97,890 saved |

When to Use TOON

✅ Best for:

• Arrays of objects with same structure (tables, lists, records)
• API responses, database results
• Large datasets sent to LLMs
• Cost optimization at scale

⚠️ Less effective for:

• Deeply nested, non-uniform data
• Small payloads (<5 items)
• Data with many unique field structures

License

MIT