Building agents that reach production systems with MCP
Source: Anthropic | 2026-04-23 URL: https://claude.com/blog/building-agents-that-reach-production-systems-with-mcp
Available Text
Building agents that reach production systems with MCP
Patterns for building effective MCP integrations: server design, OAuth with CIMD and vaults, context-efficient clients, and skills. Plus where MCP fits alongside direct API calls and CLIs for connecting agents to your systems.
Summary (Chinese)
Anthropic 发布博文，系统阐述将 Agent 接入生产系统的三种路径（直接 API 调用、CLI、MCP）及各自的适用场景。MCP 作为协议层提供通用抽象，解决了 M×N 集成问题，同时具备跨客户端（Claude、ChatGPT、Cursor 等）的可移植性。文章给出了构建高效 MCP Server 的核心设计模式：围绕意图分组工具而非端点、代码编排处理大型 API surface、标准化 OAuth CIMD 认证、支持 MCP Apps 和 Elicitation 等富语义扩展。工具搜索（tool search）可将工具定义 token 减少 85% 以上，程序化工具调用可进一步减少 37% token 使用。
Summary (English)
Anthropic's comprehensive guide on connecting AI agents to production systems, covering three integration paths (direct API calls, CLIs, MCP) with MCP as the common layer for cloud-based agents. Key patterns: intent-grouped tools, code orchestration for large APIs, OAuth CIMD auth, MCP Apps extensions. Tool search reduces tool-definition tokens by 85%+ while programmatic tool calling saves another 37%.