If you've added more than a handful of MCP servers to Claude Code or any other AI coding assistant, you've hit the wall where your context window is half-filled with JSON schemas before you write a single line of real code. It's a frustrating tax on productivity that nobody asked for—and that's exactly what mcptoon sets out to eliminate.

The Problem: Schemas Eating Your Context

MCP (Model Context Protocol) servers expose their capabilities through JSON schema definitions, and by default those schemas get shoved into the conversation context every time you invoke a tool. Add three or four MCP servers with even modest tooling surfaces, and you're already burning tokens on machine-readable metadata that humans will never read. The overhead compounds fast, especially in long coding sessions where you're invoking tools repeatedly. mcptoon takes a deceptively simple approach: instead of loading schemas into context at runtime, it keeps them stored on disk and references them only when needed. This means the AI still knows what tools are available—it just doesn't have to carry the entire schema payload in every single prompt turn. The result is a dramatic reduction in token consumption without sacrificing functionality.

Installation and Usage

The tool ships as a lightweight Python package—reported at around 50KB—and installs cleanly via pip. From there, you run MCP tools through shell commands rather than through direct context injection. It's a thin shim, but it changes the economics of how much context your AI assistant actually needs to function effectively.

Why This Matters for Developer Workflows

Token limits aren't just an abstract performance metric—they directly affect what you can accomplish in a single session. When schemas consume a significant chunk of your available context, you have less room for actual code review, refactoring decisions, or architectural discussion with the model. Any tool that recovers that space is worth evaluating seriously, particularly one that requires zero infrastructure changes and slots into existing workflows via familiar shell patterns.

Key Takeaways

  • mcptoon keeps MCP schemas on disk instead of in context, cutting token overhead by up to 99%
  • Installs via pip as a ~50KB Python package—minimal footprint, no runtime bloat
  • Tools are invoked through shell commands rather than direct context injection
  • Solves a real pain point for developers running multiple MCP servers with AI coding assistants

The Bottom Line

This is exactly the kind of boring-infrastructure win that makes a difference in day-to-day development. mcptoon doesn't do anything flashy—it just stops schemas from wasting tokens you paid for. If you're running Claude Code or similar tools with any MCP servers, this deserves a spot in your toolchain.