Tool Schema Tax

The fixed context cost paid when an agent runtime injects tool schemas into every prompt, regardless of whether those tools are used in the current turn.

Key points

  • Native MCP can be token-efficient per operation, but expensive across a full session when tool schemas are always loaded [src-041].
  • In Marco Mornati's GitHub experiment, Native MCP had the lowest per-interaction cost at 207 tokens, but the GitHub server added about 3,062 fixed schema tokens per prompt [src-041].
  • In a 20-prompt session with only two GitHub operations, the article estimates Native GitHub MCP at 61,654 tokens, with 61,240 tokens coming from repeated schema overhead [src-041].
  • The same session costs 448 tokens with raw CLI, 968 with on-demand CLI plus skill, and 892 with a gateway MCP [src-041].
  • The schema tax is most damaging for low-frequency tools that are loaded throughout long coding sessions but called only at the start or end [src-041].

Related concepts

Source references

  • [src-041] Marco Mornati — "The Future of Agentic Tooling: MCP Servers vs. CLI A Data-Driven Comparison" (2026-04-27)

Robin Cartier perspective

This page is part of Robin Cartier's working AI knowledge graph: a practical research layer for production AI, recommendation systems, experimentation, GEO, and agentic web readiness.

The useful next step is to connect this concept back to applied product leadership and operating models.

Recommended next

Keep reading from this thread

From 500 indexed pages and articles.

  1. Wiki concept Gateway MCP Pattern Tool-access architecture where the agent sees one small MCP router while backend service schemas are loaded or invoked lazily. Related by schema
  2. Wiki concept G/N Ratio Tool Selection Decision framework for choosing between CLI, on-demand skill files, Gateway MCP, and Native MCP by estimating how often a service is actually Related by schema
  3. Wiki concept On-Demand Skill Files Pattern for keeping service-specific CLI/API guidance out of the always-loaded project context until the agent actually needs that service. Related by 041