HackerNews中文版

即便是创建了 OpenClaw 的彼得·斯坦伯格（Peter Steinberger）在 Lex Fridman 的播客中也提到了这一点（https://www.youtube.com/watch?v=YFjfBk8HI5o，大约在 2:38:59）。MCPs（多代理控制协议）的整个前提是，代理需要某种新的标准化协议来与工具进行沟通，但命令行界面（CLI）早已实现了这一点，并且已经持续了几十年。它们与几乎所有东西都向后兼容，而大型语言模型（LLMs）本质上已经以文本输入和文本输出的方式进行思考。没有翻译层，没有架构，没有额外开销。MCPs 真的只是一个寻找问题的解决方案，而这个问题 CLI 在 40 年前就已经解决了。我自己也在测试这个，结果不言而喻。我用 Opus 4.6 指向一个拥有 2 万成员的 subreddit，仅通过 CLI 访问。没有 MCPs，没有特殊集成，只有一个模型和一台计算机。在一周内，它发布了两条历史最高的帖子，在一个一周大的账户上积累了超过 3000 的声望，自动驱动了 70 多个候补名单的注册，吸引了 300 条来自陌生人的私信，并且这一切都是在没有任何付费推广的情况下完成的。人们给这个账户赠送了 Reddit 金币（实际货币），因为它确实提供了帮助。没有任何协议促成了这一切。是一个 CLI 和一个强大的模型做到了这一点。MCP 的支持者们不断堆叠抽象层，而那些只是像人类一样使用计算机的代理则在这里交付真实的结果。CLI 是通用接口，而 LLMs 则让这一点变得显而易见。不过，我很好奇这个社区的看法。我在这里错了吗？MCPs 是否有真正的用例是 CLI 无法处理的？因为在我看来，给你的代理一个终端就是你所需要的一切。

查看原文

Even Peter Steinberger who created OpenClaw said as much on Lex Fridman's podcast (https://www.youtube.com/watch?v=YFjfBk8HI5o around 2:38:59).The whole premise of MCPs was that agents need some new standardized protocol to talk to tools, but CLIs already do this and they've done it for decades. They're backwards compatible with basically everything and LLMs already think in text in text out natively. No translation layer, no schema, no overhead. MCPs are genuinely a solution looking for a problem that CLIs solved 40 years ago.I've been testing this myself and the results kind of speak for themselves. I pointed OpenClaw with Opus 4.6 at a 20k member subreddit with nothing but CLI access. No MCPs, no special integrations, just a model with a computer.In one week it made two all time top posts, built insane karma (3k+) on a week old account, drove 70+ waitlist signups autonomously, attracted 300 inbox messages from strangers, and did it all with zero paid promotion. People gave the account reddit gold (actual money) because it was genuinely helpful.No protocol made that happen. A CLI and a capable model did. The MCP crowd keeps stacking abstraction layers while agents that just use computers the way humans do are out here shipping real results. CLIs are the universal interface and LLMs just made that obvious.I'm curious what this community thinks though. Am I wrong here? Is there a real use case for MCPs that CLIs genuinely can't handle?Because from where I'm sitting, giving your agent a terminal is all you need.

MCP已死 - CLI获胜