MCP 如何为 LLM 解决工具发现问题

Оригинальный заголовок: MCP Solves Tool Discovery for LLMs

Заголовок и краткое изложение на выбранном языке ожидают перевода.

Краткий обзор ИИ

MCP 通过提供一份实时、机器可读的工具目录，为 Claude Code、OpenAI Codex 等编程智能体解决工具发现问题——目录包含工具名称、描述、输入 schema 和示例调用。CLI 工具对编程智能体不可读，而 MCP 内置了自动提示模型的机制，让模型无需猜测或搜寻工具。作者认为，智能体优先的开发意味着从工具编写之初就要面向模型提示，这正是 MCP 所内建的。

Полный текст

Coding agents like Claude Code and OpenAI Codex don’t struggle with terminals; they struggle to discover tools that exist. If a command isn’t in the model’s training set or post-training hill climbing, they won’t even know to try it. MCP fixes discovery by giving models a live, machine-readable catalog of tools with names, descriptions, input schemas, and example calls. MCP gives the models tokens.

A 3D-rendered, diorama-style illustration shows a small, rounded robot with glossy black eyes sorting through a wooden crate filled with hand tools. The robot holds a hammer in one hand, with other tools like wrenches and a screwdriver neatly arranged in the crate. The scene features soft, neutral studio lighting and warm, natural tones, creating a minimal, premium look.

Unix pipes and man pages gave humans both composition and discovery. In agentic systems, LLMs provide composition and MCP gives LLMs discoverability + affordances.

MCP is the universal plugin interface, but LLM behavior is fundamentally driven by, and constrained by, input tokens. MCP offers something that CLIs ambiently available in $PATH (or training data) never could: a built-in mechanism to automatically prompt the model so it doesn’t have to guess or hunt for tools.

The CLI your developer productivity team built to accelerate human developers is illegible to a coding agent. Agent-first development means prompting the model from the start — and MCP bakes that into tool authorship.

Источник: Ryan Lopopolo · hyperbo.la