📄 中文摘要
MCP（模型上下文协议）在六个月前被视为 AI 工具的未来，得到了各大 AI 实验室的支持。然而，Perplexity 在将其投入生产后，发现了一个令人震惊的事实：其内部的后期分析显示，81% 的 MCP 上下文预算被协议开销消耗，而非实际工具内容。这并不是配置问题，而是规范按设计运行的结果。具体而言，一个包含 12 个工具的典型 MCP 工具清单，仅用于描述可用工具就消耗了约 4200 个令牌，这严重影响了其令牌预算。
生产中的 MCP：Perplexity 的首席技术官为何选择离开（以及数据所揭示的真相）
📄 English Summary
"MCP in Production: Why Perplexity's CTO Walked Away (And What the Data Says)"
Six months ago, Model Context Protocol (MCP) was hailed as the future of AI tooling, receiving endorsements from major AI labs. However, after Perplexity deployed it in production, the internal post-mortem revealed a startling statistic: 81% of their MCP context budget was consumed by protocol overhead rather than actual tool content. This was not a configuration issue, but rather the specification functioning as designed. For instance, a typical MCP tool manifest with 12 tools consumed approximately 4,200 tokens just to describe the available tools, significantly impacting their token budget.