A developer has slashed MCP costs by 94% using a simple CLI conversion approach. The innovation targets a hidden expense in AI agent workflows: the instruction manual. While API calls remain efficient, MCP servers dump entire tool catalogs as JSON Schema into conversations, creating massive token overhead before agents can perform any useful work.

By converting MCP servers to lightweight CLIs using CLIHub, the same tools and OAuth functionality are preserved while dramatically reducing token usage. The approach uses lazy loading - agents discover tool details only when needed. For a typical setup with 6 servers and 14 tools each, MCP loads 15,540 tokens upfront versus just 300 tokens for CLI's skill listing. Tool calls also benefit: MCP's pre-loaded calls use 30 tokens while CLI's discovery approach totals around 910 tokens for the same operation.

Anthropic's Tool Search achieves similar goals but remains more expensive and limited to Anthropic models. CLI works universally and costs less overall. The developer open-sourced the converter, making it a one-command process to transform MCP servers into CLIs. This practical solution addresses a widespread inefficiency affecting every AI agent using MCP, potentially saving significant costs across the ecosystem.