mcptoon — CLI that cuts MCP tool discovery from 2034 to 62 tokens (97% reduction)

Talking Shop
1

I use MCP servers with Claude Code daily. Last week I counted how many tokens the tool listing eats.

5 servers, 96 tools. The JSON listing: 2,034 tokens. Before I type anything. Then 20 tool calls, each wrapped in {"content":[{"type":"text","text":"..."}]} — another 40K tokens of brackets, quotes, and repeated {"type":"object","properties": declarations.

So I built mcptoon — a CLI that outputs TOON (Token-Optimized Object Notation) instead of JSON.

What TOON does

  • {"name":"search","count":3} → name:search|count:3
  • [1, 2, 3] → 1 2 3
  • true/false → T/F
  • null → ∅

Measured results

Operation JSON tokens TOON tokens Savings
Tool discovery (96 tools) 2,034 62 97%
Tool result (structured data) 812 354 56%
Real session (5 servers, 20 calls) 47,200 8,100 83%

It’s a CLI, not a library. Any agent that runs shell commands can use it. Zero deps, 50KB, Python 3.10+, Apache 2.0.

Tested TOON parsing with Claude, GPT-4, and Gemini — all three parse it correctly.

pip install mcptoon

GitHub: GitHub - activeing123/mcptoon: Token-efficient MCP CLI client. 97% less tokens on tool discovery, 40-60% on results. Zero deps. Cross-platform. Works with every AI agent. · GitHub

Curious if others have measured their MCP token overhead. My numbers are in the README, would be interested to compare.

1 Like