How the design of an MCP server affects agent token usage

Описание и применение

Добавление композитных операций в MCP сократило количество вызовов и токенов, необходимых агенту для редактирования Word‑документов, улучшив эффективность без потери корректности.

Задача

MCP servers often require agents to perform many small tool calls, increasing output tokens and protocol overhead when editing Word documents.

Результат

Composite MCP operations cut MCP calls by ~29% and reduced output tokens by 4–16% across models, improving efficiency without sacrificing correctness.

Рабочий процесс

• Identify frequent multi‑call sequences in agent traces
• Create composite operations that combine those calls
• Add the composites to the MCP tool surface without removing primitives
• Run controlled experiments comparing shortcut vs. no‑shortcut surfaces
• Measure MCP calls, output tokens, latency, and correctness
• Analyze aggregate results and remove ineffective shortcuts

Доказательства

Total number of MCP calls fell from 501 to 358—a reduction of approximately 29%

Источник

Median of the seven task‑level shortcut/no‑shortcut output‑token ratios was 0.85x to 0.96x

Источник