📄 中文摘要
2026年3月，AI代理领域最热门的话题并不是哪个模型更智能，而是一个看似无聊的架构问题：代理应该如何调用外部工具。对此，存在三个主要阵营。MCP阵营（模型上下文协议）提出了一种开放标准，旨在通过统一的JSON-RPC协议封装服务接口，使代理能够在单次集成后跨平台调用多个工具。CLI阵营则主张让代理直接运行shell命令，如`git log`和`gh pr list`，以实现更灵活的操作。这场争论反映了对AI代理如何高效利用外部工具的不同看法。
CLI vs MCP vs 技能：整个争论问错了问题
出处: CLI vs MCP vs Skills: The Whole Debate Is Asking the Wrong Question
发布: 2026年3月27日
📄 English Summary
CLI vs MCP vs Skills: The Whole Debate Is Asking the Wrong Question
In March 2026, the most debated topic in the AI agent realm is not about which model is smarter, but rather a seemingly mundane architectural question: How should an agent call external tools? Three main camps have emerged. The MCP Camp (Model Context Protocol) advocates for an open standard launched by Anthropic in late 2024, which wraps service interfaces in a unified JSON-RPC protocol, allowing agents to call multiple tools across platforms after a single integration. In contrast, the CLI Camp argues for enabling agents to run shell commands like `git log` and `gh pr list`, promoting greater flexibility. This debate highlights differing perspectives on how AI agents can efficiently leverage external tools.