toon-mcp-server

MCP server for token-efficient data encoding in LLM context windows.

v2 uses LEAN format (LLM-Efficient Adaptive Notation) as the primary encoding, with TOON available for comparison in stats.

Why?

JSON wastes tokens on syntax — braces, brackets, repeated key names, quotes. LEAN saves ~49% tokens on average by declaring field names once in tab-delimited tables, using dot-flattened keys, indentation instead of braces, and single-char literals (T/F/_).

Tools

• lean_encode — JSON value -> LEAN
• lean_decode — LEAN -> JSON value
• lean_stats — compare token counts: JSON vs LEAN vs TOON
• lean_format_response — picks whichever format is smaller (LEAN or compact JSON), adds metadata

All tools accept any JSON value at root. Input validated with Zod.

Setup

npm install && npm run build

Claude Desktop (claude_desktop_config.json):

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

Claude Code:

claude mcp add toon -- node /absolute/path/to/toon-mcp-server/dist/index.js

Limitations

Keys must match [\w][\w-]*. No dots, spaces, slashes, colons. encode() throws on bad keys.

Strings round-trip correctly including empty strings, scalar-lookalikes ("null", "42"), whitespace, backslashes, newlines. Two escape conventions:

• Scalar context (key:value): backslash escaping (\\, \", \n)
• Cell context (tabular rows): RFC 4180 doubling ("") + \n

Empty root object encodes as {}.

Files

index.ts      MCP server, 4 tools
lean.ts       LEAN encoder/decoder + token estimation
toon.ts       TOON encoder/decoder (used in stats comparison)
test.mjs      tests
dist/         compiled output

License

MIT