DEV Community

gentic news
gentic news

Posted on • Originally published at gentic.news

Cut MCP Token Waste 99%: The 50KB CLI That Keeps Schemas Out of Context

#ai #programming #tech #product

mcptoon CLI slashes MCP token overhead by up to 99% by keeping schemas on disk—install with pip and run tools via shell commands.

MCP Token Bloat: The Silent Context Killer

How to Cut Your MCP Token Usage by 97% in 3 Minutes | by Piotr | Medium

If you've added more than a few MCP servers to Claude Code, you've felt it: the context window fills with JSON schemas before you even start. The MCP community itself says "MCP is unusable past 5 servers," and Latent Space's analysis confirms a "scaling cliff around 20-30 tools."

Why? Every MCP server you connect pushes its tool definitions into your agent's context as verbose JSON. Browser tools like puppeteer or playwright alone can add 50,000-100,000+ tokens of {"type":"object","properties":...} noise. That's context you're paying for but never using.

The Fix: mcptoon CLI

mcptoon (GitHub: activeing123/mcptoon) is a 50KB Python CLI that flips the model: instead of Claude Code connecting to MCP servers as a client, your agent runs mcptoon commands directly. Schemas live on disk in ~/.mcptoon/config.json—they never enter context. Only the compact output you request does.

Measured Results (255 tools, 5 formats)

Tools JSON Standard TOON mcptoon SLIM Compact
50 17,790 8,776 (-51%) 6,981 (-61%) 1,203 (-93%) 117 (-99%)
255 90,804 44,863 (-51%) 35,735 (-61%) 6,174 (-93%) 117 (-100%)

With 255 tools, JSON costs 90,804 tokens. mcptoon's compact mode: 117 tokens. That's a 99.87% reduction—freeing your context for actual work.

TOON vs JSON

JSON (287 tokens for one tool):

[{"name":"search_web","description":"Search the web","inputSchema":{"type":"object","properties":{"query":{"type":"string"},"num_results":{"type":"number","default":5}},"required":["query"]}}]
Enter fullscreen mode Exit fullscreen mode

TOON (5 tokens):

search_web fetch_url
Enter fullscreen mode Exit fullscreen mode

Simon Willison nailed it: "JSON is the least token-efficient format possible for structured data sent to an LLM." TOON strips away everything but the essentials.

Setup: 3 Commands

Cut your Claude Code’s token consumption by 90% 🤯 Someone built RTK, a ...

pip install mcptoon
mcptoon init
mcptoon add fetch --stdio npx -y @modelcontextprotocol/server-fetch
mcptoon manifest --compact  # → all tool names, 117 tokens
Enter fullscreen mode Exit fullscreen mode

Your agent can even install its own tools:

mcptoon add github --stdio npx -y @modelcontextprotocol/server-github
Enter fullscreen mode Exit fullscreen mode

Works with Claude Code, Cursor, Codex, OpenCode, CatPaw—any agent that runs shell commands. Zero dependencies, Apache 2.0.

Why This Matters for Claude Code Users

  • Context death solved: Keep 10+ servers without eating your window.
  • Config hell gone: No more hand-editing claude_desktop_config.json—one config file, all tools.
  • Agent self-service: Your agent can add tools on the fly via mcptoon add.
  • Reconfigure once: Same config works across all agents—no per-agent redo.
  • Stop paying for JSON garbage: Every MCP result in JSON wastes ~80 tokens for 6 tokens of data. Over 200 tool calls, that's 15,000 tokens of syntax waste.

Try It Now

  1. Install: pip install mcptoon
  2. Init: mcptoon init
  3. Add your most-used servers (fetch, github, etc.)
  4. Ask Claude Code to run mcptoon manifest --compact and see the token savings instantly.

Your context window is precious. Stop letting MCP schemas eat it alive.

Source: dev.to

Originally published on gentic.news

Top comments (0)