Проектирование MCP-серверов для AI-агентов

★ 7.1 · general

mcp-for-agents is a Claude Code skill that helps design and review MCP servers so AI agents can use them reliably. It covers outcome-oriented tool design — combining backend operations server-side so the agent makes one call instead of three — alongside flat, constrained parameters with enums and defaults that reduce hallucination of missing keys. The skill details how to write descriptions that trigger correct tool selection, citing research showing 97% of tool descriptions across 103 MCP servers have quality deficiencies. Error handling, token efficiency, and composable outputs are also addressed: tool input schemas account for 60–80% of total MCP token usage, and keeping a server to 5–15 tools is the practical ceiling for reliable selection. Use it when building or reviewing an MCP server, adding tools, or debugging an agent that picks the wrong tool or produces bloated responses.