MCP Design Pitfall: Exposing APIs As-Is Wastes Context Window
The article provides concrete evidence that naive MCP server design undermines AI agent accuracy, with measurable context window costs that developers must account for.
Reporting from 1 source: ASCII.jp.
A new article from CData, re-edited on ASCII.jp, warns that directly mapping existing API endpoints to MCP tools degrades AI agent performance. Each tool consumes 200-600 tokens, and 150 tools can eat 25-30% of a 200k context window before any task data. Cursor and GitHub Copilot enforce tool limits of 40 and 128 respectively, based on measured degradation. The piece advocates for scoped tool design to reduce context tax.
Each MCP tool adds 200 to 600 tokens to the context window just for its name, description, and parameter schema. With five servers each offering 30 tools, that is 150 tools occupying a quarter to a third of a 200,000-token context before any user message or retrieved data arrives. Cursor caps MCP tools at 40, and GitHub Copilot enforces a hard limit of 128-both thresholds derived from observed performance drops beyond those counts.
The article, originally published on the CData Blog and re-edited on ASCII.jp, identifies the root cause: developers wrap existing REST endpoints as MCP tools without scoping, carrying over API redundancy into the LLM's context. Fine-grained CRUD operations that could be combined into one tool instead multiply the overhead. The fix is scoped tool design that minimizes the manifest size per request.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.