最近对 MCP 服务器的分析发现，大约 10.1% 的服务器确实存在故障，这与广泛流传的“一半服务器已失效”的说法相反。较高的数字可能是由于将需要 API 密钥的服务器错误地归类为非功能性。除了简单的正常运行时间，一个更重要的问题是模式漂移，即工具合同在 LLM 不知情的情况下发生变化，导致静默故障。这种漂移会导致代理程序出现故障，表现为 AI 错误，而不是底层服务问题。建议开发人员实施严格的服务器活动性和模式一致性检查，以确保可靠运行。 AI
影响 强调了依赖外部工具的 LLM 代理的关键可靠性问题，并强调了对稳健的错误处理和模式验证的需求。
排序理由 对特定技术协议 (MCP) 及其故障模式的分析。
AI 生成摘要 · Google Gemini · 来自 3 个来源。 我们如何撰写摘要 →
<p>There is a number going around that roughly half of all remote MCP servers are dead.</p> <p>I had repeated it myself, in the README of a tool I published. I could not find where it came from, so I measured it.</p> <p>The answer is that about <strong>one in ten</strong> is actu…
<p>There is a number going around that roughly half of all remote MCP servers are dead.</p> <p>I had repeated it myself, in the README of a tool I published. I could not find where it came from, so I measured it.</p> <p>The answer is that about <strong>one in ten</strong> is actu…
<p>Every few weeks, one of my MCP servers starts returning garbage. Not errors — worse. Silent corrupted data. Responses that look correct until you look at the timestamps, or the IDs, or the file paths.</p> <p>I've been running MCP servers in production for about a year now. Aft…