[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"pack-detail-mcp-monitoring-logs-zh":3,"seo:pack:mcp-monitoring-logs:zh":92},{"code":4,"message":5,"data":6},200,"操作成功",{"pack":7},{"slug":8,"icon":9,"tone":10,"status":11,"status_label":12,"title":13,"description":14,"items":15,"install_cmd":91},"mcp-monitoring-logs","📊","#0891B2","stable","稳定","MCP 监控 + 日志直查","9 个 MCP server 和 subagent，把 Prometheus \u002F Grafana \u002F Sentry \u002F Datadog \u002F SigNoz 变成 AI agent 真能直接查的东西。日志、指标、trace、dashboard、告警全串起来 — 凌晨 02:47 checkout 为啥炸了，不用人 grep。",[16,28,35,42,50,57,67,77,84],{"id":17,"uuid":18,"slug":19,"title":20,"description":21,"author_name":22,"view_count":23,"vote_count":24,"lang_type":25,"type":26,"type_label":27},3608,"818380f9-674d-5217-88ab-f393ff99a247","signoz-mcp-server-query-traces-logs-alerts","SigNoz MCP Server — Query Traces, Logs & Alerts","SigNoz MCP Server connects MCP clients to your SigNoz instance: query traces\u002Flogs, inspect alerts, and automate observability workflows using an API key.","MCP Hub",256,0,"en","mcp","MCP",{"id":29,"uuid":30,"slug":31,"title":32,"description":33,"author_name":22,"view_count":34,"vote_count":24,"lang_type":25,"type":26,"type_label":27},827,"655bef8a-41bb-4eda-8f5b-29d1d4cb8c74","axiom-mcp-log-search-analytics-ai-agents-655bef8a","Axiom MCP — Log Search and Analytics for AI Agents","MCP server that gives AI agents access to Axiom log analytics. Query logs, traces, and metrics through natural language for AI-powered observability and incident response.",264,{"id":36,"uuid":37,"slug":38,"title":39,"description":40,"author_name":22,"view_count":41,"vote_count":24,"lang_type":25,"type":26,"type_label":27},3191,"0be32c8b-2ad9-47f2-aa64-26f9a7f6f2c3","grafana-mcp-dashboards-alerts-oncall-tools","Grafana MCP — Dashboards, Alerts & OnCall Tools","Grafana MCP server connects your MCP client to Grafana so agents can search dashboards, query panels, and automate ops tasks with a service account token.",39,{"id":43,"uuid":44,"slug":45,"title":46,"description":47,"author_name":48,"view_count":49,"vote_count":24,"lang_type":25,"type":26,"type_label":27},2853,"b309576b-0970-46f2-b0f8-a2f1af76eeb1","datadog-mcp-server-query-metrics-and-logs-from-ai-agents","Datadog MCP Server — Query Metrics and Logs from AI Agents","Community Datadog MCP server lets Claude \u002F Cursor query metrics, logs, traces, monitors in natural language. SRE copilots, on-call triage.","Datadog",212,{"id":51,"uuid":52,"slug":53,"title":54,"description":55,"author_name":22,"view_count":56,"vote_count":24,"lang_type":25,"type":26,"type_label":27},665,"a739e813-e8fa-4285-8634-55aa447dd71a","sentry-mcp-error-monitoring-server-ai-agents-a739e813","Sentry MCP — Error Monitoring Server for AI Agents","MCP server that connects AI agents to Sentry for real-time error monitoring. Query issues, analyze stack traces, track regressions, and resolve bugs with full crash context. 2,000+ stars.",314,{"id":58,"uuid":59,"slug":60,"title":61,"description":62,"author_name":63,"view_count":64,"vote_count":24,"lang_type":25,"type":65,"type_label":66},2276,"676b8063-2e21-49ce-89eb-020bcc40cb47","sentry-errors-auto-triage-subagent-for-sentry-676b8063","sentry-errors — Auto-Triage Subagent for Sentry","Open-source Claude Code subagent that pulls recent Sentry errors via MCP, groups by component, and suggests fix priorities. Inspired by Boris Cherny.","Skill Factory",232,"skill","Skill",{"id":68,"uuid":69,"slug":70,"title":71,"description":72,"author_name":73,"view_count":74,"vote_count":24,"lang_type":25,"type":75,"type_label":76},3546,"2c1c7883-fd56-5cb3-91fe-6f9d953193f3","pup-datadog-cli-companion-for-ai-agents","Pup — Datadog CLI Companion for AI Agents","Pup is an Apache-2.0 Datadog CLI with 200+ commands across 33+ products, so agents can query logs, metrics, RUM, and security data via OAuth login.","Script Depot",145,"script","Script",{"id":78,"uuid":79,"slug":80,"title":81,"description":82,"author_name":22,"view_count":83,"vote_count":24,"lang_type":25,"type":26,"type_label":27},3444,"71f97e34-fa9c-5c0b-8c21-69d6570cb21f","langfuse-mcp-query-langfuse-traces-via-mcp","langfuse-mcp — Query Langfuse Traces via MCP","Connect Langfuse observability to Claude Code\u002FCodex via MCP: fetch traces, prompts, and datasets (37 tools). Works with Langfuse Cloud or self-hosted.",225,{"id":85,"uuid":86,"slug":87,"title":88,"description":89,"author_name":73,"view_count":90,"vote_count":24,"lang_type":25,"type":65,"type_label":66},3335,"a86f3430-eb78-50ab-bebe-6eef4f53ea4a","monoscope-llm-query-for-logs-traces-metrics","Monoscope — LLM Query for Logs\u002FTraces\u002FMetrics","Monoscope stores logs\u002Ftraces\u002Fmetrics in S3-compatible buckets and lets you explore them with natural-language queries plus a CLI and self-hosted UI.",177,"tokrepo install pack\u002Fmcp-monitoring-logs",{"pageType":93,"pageKey":8,"locale":94,"title":95,"metaDescription":96,"h1":97,"tldr":98,"bodyMarkdown":99,"faq":100,"schema":116,"internalLinks":122,"citations":135,"wordCount":148,"generatedAt":149},"pack","zh","MCP 监控 + 日志直查 — 给 Prometheus \u002F Grafana \u002F Sentry \u002F Datadog 接上 AI agent 的 9 件套","SigNoz MCP \u002F Axiom MCP \u002F Grafana MCP \u002F Datadog MCP \u002F Sentry MCP \u002F sentry-errors 自动三角化 subagent \u002F Pup (Datadog CLI) \u002F langfuse-mcp \u002F Monoscope —— 把你已经在用的监控栈接到 AI agent。日志 → 指标 → trace → 告警 → dashboard，按推荐安装顺序。","MCP 监控 + 日志直查 — 你可观测性栈的 agent 接入层","9 个 MCP 按推荐安装顺序：先日志查询 MCP（SigNoz \u002F Axiom），再指标 + dashboard MCP（Grafana \u002F Datadog），再错误 + 告警 MCP（Sentry + sentry-errors 三角化 subagent），再 CLI 兜底（Pup），最后 LLM trace MCP（langfuse-mcp \u002F Monoscope）。目标：agent 看一眼 Sentry → 拉 SigNoz trace → MCP 查 Loki → 写好事故摘要 —— 全程你不用打开 Grafana。","## 这个 pack 是什么 —— 以及不是什么\n\n这个 pack 不是一套监控栈。**你已经有一套了**。Prometheus 在抓指标、Grafana 在渲图、Sentry 在聚合异常，Loki 或 SigNoz 也许在吃日志。pack 是叠在这套栈上的 **agent 接入层** —— 就是那些 MCP server 和 subagent，让 Claude \u002F Cursor \u002F ChatGPT 或任何 MCP 客户端能直接用自然语言查 trace、日志、告警、dashboard，不需要人去点。\n\n如果你想要底层日志流水线（winston \u002F Fluent Bit \u002F Loki \u002F ClickHouse \u002F lnav），看 [log-analysis-search pack](\u002Fzh\u002Ftopics)。如果你想要更广的 deploy + monitor 栈（Vercel \u002F Uptime Kuma \u002F OpenTelemetry Collector \u002F Alertmanager），看 [deploy-monitor-observability pack](\u002Fzh\u002Ftopics)。本 pack 假设上面那些你已经有了，专门补齐 MCP 接入层，让 agent 真能用起来。\n\n为啥这事重要：大部分团队已经在花钱（或自建）了足够多的可观测工具。**他们缺的是一条 on-call 流程** —— 告警来了能自动三角化、SRE 问「02:47 改了啥」能一句话拿到答案、上线出事故时摘要在人打开 dashboard 之前就写好。MCP 接对了就是干这个的。\n\n每一个选品都是**开源、或者有可自建的开源 MCP server**。有几个包的是商业后端（Datadog、Sentry SaaS、Axiom），但 MCP server 本身是你能读源码再跑的开源代码。**不收黑箱 agent SDK**。\n\n## 推荐安装顺序\n\n1. **SigNoz MCP Server** — 日志 + trace + 告警查询 MCP。从这里开始，因为 SigNoz 是少数开源后端里原生把日志\u002Ftrace\u002F指标统一进一个存储的。一旦 agent 能查 SigNoz，它就能回答「过去一小时最慢的接口」「这个错首次出现是什么时候」「现在哪些告警在烧」——不用三个工具来回跳。如果你不用 SigNoz，跳到第 2 步。\n2. **Axiom MCP — Log Search and Analytics for AI Agents** — 云端日志搜索 MCP。当你日志在托管存储里、查询用 APL（Axiom Processing Language）而不是 LogQL 时的替代方案。和 SigNoz MCP 是同一种活，后端不同。看你日志在哪里二选一，两个都装就是浪费。\n3. **Grafana MCP — Dashboards, Alerts & OnCall Tools** — dashboard MCP，整个 pack 的支柱。Grafana 的 MCP 把面板数据、告警规则、OnCall 排班、dashboard 搜索全暴露出来。装完之后，agent 能拉你手动打开的同一张图，读底层 PromQL\u002FLogQL 表达式，并对它做推理。不装这个，agent 对 Grafana 渲染的一切就是瞎的。\n4. **Datadog MCP Server — Query Metrics and Logs from AI Agents** — Datadog 用户的对应方案。和第 1+3 步合起来是一个职能，但走 Datadog 的指标\u002F日志\u002FAPM。**默认只读是底线安全姿势**，开放给自主 agent 之前一定再核一遍。\n5. **Sentry MCP — Error Monitoring Server for AI Agents** — 错误 MCP。Sentry 官方 server，能返回 issue 列表、stack trace、回归状态、release 健康度。凌晨 3 点 agent 三角化告警从这里开始 —— 「这是新错还是已知回归」是一个工具一句话就能答的问题。\n6. **sentry-errors — Auto-Triage Subagent for Sentry** — 告警工作流 agent。这是叠在 Sentry MCP 之上的层：一个 Claude subagent，告警一响它就醒，调 Sentry MCP 拉 issue，调 SigNoz\u002FGrafana MCP 拉上下文，写出结构化的三角化笔记。你可以自己写，但这个已经验证过；fork 就好。\n7. **Pup — Datadog CLI Companion for AI Agents** — CLI 桥。当 agent 的 MCP toolset 表达不了某个查询（或你不信任 MCP 暴露写操作）时，Pup 给 agent 一个沙箱化的 Datadog CLI。默认只读 flag，配合 audit log。\n8. **langfuse-mcp — Query Langfuse Traces via MCP** — LLM trace MCP。分类不同：它查的不是你应用的请求 trace，而是 Langfuse 里的 **LLM trace** —— prompt、response、tool call、cost。当 on-call 问题是「02:47 agent 为啥答错了」而不是「02:47 checkout 为啥炸了」时就是它。**线上跑 LLM 功能的话必装**。\n9. **Monoscope — LLM Query for Logs\u002FTraces\u002FMetrics** — 统一自然语言查询层。架在多个后端（日志、trace、指标）前面，让 agent 或人能直接问「过去 30 分钟 checkout 模块 p99 > 500ms 的错误」——不用先决定走哪个工具。当你逐个 MCP 调嫌烦的时候上它。\n\n## 流水线怎么协同\n\n```\n   [ 告警触发 ]\n         │\n         ▼\n   sentry-errors subagent           ← 唤醒、编排\n         │\n         ├──▶ Sentry MCP             (issue、stack trace、是否回归？)\n         ├──▶ SigNoz \u002F Axiom MCP     (时间戳前后的日志)\n         ├──▶ Grafana MCP            (面板状态、告警规则、OnCall 名单)\n         ├──▶ Datadog MCP \u002F Pup CLI  (Datadog 用户的指标查询)\n         ├──▶ langfuse-mcp           (路径里有 AI 功能时拉 LLM trace)\n         │\n         ▼\n   Monoscope (可选)                  ← 统一 NL 查询、fan-out\n         │\n         ▼\n   [ 结构化三角化笔记 → 工单 \u002F Slack \u002F 电话 ]\n```\n\n形状是有讲究的：**subagent 在最上面、MCP server 在中间、后端存储在最底**。每个 MCP 都是薄壳，价值在组合。单调用一次 SigNoz MCP 很有意思；sentry-errors 在 4 个 MCP 上 fan-out 然后写一段话，那才是真正的解锁。\n\n如果你从零开始：第 4 步（Datadog）和第 7 步（Pup）你不在 Datadog 就跳。第 8 步（langfuse-mcp）不跑 LLM 功能就跳。**最小可用组合是 SigNoz MCP + Grafana MCP + Sentry MCP + sentry-errors subagent —— 4 个选品**，agent 已经能回答凌晨 3 点的大部分问题。\n\n## 你会遇到的取舍\n\n- **MCP vs 自定义 function-calling** — 这里每一个 MCP server 都可以用 OpenAI function-calling 重写一遍对同一个后端 API。MCP 胜在你有不止一个 agent runtime 时（Claude Desktop、Cursor、ChatGPT 自定义 GPT、内部 agent）—— 一次写 MCP，所有客户端复用。Function-calling 胜在你只有一个定制 agent + 一个客户端。\n- **只读 vs 读写 scope** — 这里的每一个监控 MCP 都能配成读写（ack 告警、静默告警、改 dashboard）。on-call 三角化，**只读是唯一合理默认**。读写是每个 server 单独的决定，必须配 audit log，并强制走人审批。\n- **开源 MCP vs 厂商 MCP** — Grafana MCP 和 Sentry MCP 是 Grafana Labs \u002F Sentry 一方维护的。SigNoz MCP 和 Axiom MCP 也是一方。Datadog MCP 和 Pup 是社区针对 Datadog API 维护的。一方更稳定；社区在边缘特性上迭代更快。**部署前看一眼 maintainer**。\n- **sentry-errors vs 自己写 subagent** — sentry-errors 是一套有自己态度的三角化流程。如果你事故 playbook 不一样（你先 page 再三角化），用起来会别扭。fork 就行，**价值在 fan-out 模式，不在具体那段 prompt**。\n- **Monoscope vs 单独 MCP server** — Monoscope 是统一器。你 MCP 不到 4 个之前不需要它。当 fan-out 延迟或 token 成本真的成问题时再加。\n\n## 常见踩坑\n\n- **默认暴露写权限** — 几乎所有 MCP server 文档都先举写操作示例（因为更炫）。监控\u002F可观测 MCP 这种地方，**agent 不该有静默告警、ack incident、改 dashboard 的能力**，除非走一条明确的人审批通道。每个 MCP 配置必审。\n- **大日志查询的 agent token 成本** — agent 问「过去 24 小时所有错误」走 SigNoz MCP 能拉几 MB 进上下文。在 MCP server 配置里设响应大小上限、超过 N 行就拒并要求 agent 加过滤条件。\n- **一次 MCP 调用混查日志和指标** — agent 会尝试。大部分后端一次只回答一件事是擅长的。**纪律编进 system prompt**：先问 SigNoz MCP 拿 trace ID，再问 Grafana MCP 拿指标面板，不要一句问俩。\n- **MCP server 之间没相关性** — `trace_id` 是胶水。如果你的日志、trace、指标没共享 `trace_id`，agent 的 fan-out 拉到四件不相关的东西然后幻觉出连接。**修 instrumentation 优先于修 agent**。\n- **MCP server 和生产跑在同一台 VM** — 只读 MCP 在 agent 突发使用下也占内存和 CPU。**MCP 跑在独立小 VM**，和可观测后端隔离，免得失控 agent 把日志存储 OOM 了。\n- **sentry-errors 没限流** — 告警风暴时（上线回归、基础设施事件）subagent 每条告警都醒。在告警源做去重窗口（Alertmanager \u002F Sentry 规则），或限流 subagent 调用次数。**别为同一个根因付 Anthropic API 500 张三角化笔记的钱**。",[101,104,107,110,113],{"q":102,"a":103},"这 9 个必须全装吗？最小可用组合是什么？","最小是 4 个：SigNoz MCP（或日志在 Axiom 就上 Axiom MCP）+ Grafana MCP + Sentry MCP + sentry-errors 三角化 subagent。这 4 个覆盖日志、dashboard\u002F告警、错误、编排层。Datadog MCP 和 Pup 只有你公司在 Datadog 上才加。langfuse-mcp 只有你线上跑 LLM 功能需要 LLM trace 才加。Monoscope 只有 fan-out 在 4 个以上 MCP 上 token 成本超标才加。大多数团队最终落在 5-6 个选品。",{"q":105,"a":106},"这个 pack 和 log-analysis-search、deploy-monitor-observability 有啥区别？","那两个 pack 是底层可观测栈：log-analysis-search 覆盖 winston\u002FLoguru\u002FFluent Bit\u002FLoki\u002FClickHouse\u002Flnav（日志怎么存、怎么读），deploy-monitor-observability 覆盖 Prometheus\u002FGrafana\u002FUptime Kuma\u002FAlertmanager（栈怎么上线）。**本 pack 是叠在这两个栈上的那一层**，让 AI agent 能查那两个 pack 搭好的东西。是故意做的互补 —— 先装那两个之一，再装这个。",{"q":108,"a":109},"MCP server 该只读还是读写？","**本 pack 里所有 MCP 都默认只读**。本职是观察、不是修改。读写 MCP（ack 告警、静默告警、改 dashboard）是每个 server 单独决策，必须配 audit log，并要走 agent 工作流里的人审批一步。**风险不对称** —— agent 读 dashboard vs agent 静默一条真告警 —— 让只读成为没有歧义的默认。",{"q":111,"a":112},"MCP server 和 sentry-errors 这种 MCP subagent 有什么区别？","MCP server（Sentry MCP \u002F SigNoz MCP \u002F Grafana MCP）把一个后端的 API 暴露成 agent 可调用的工具集。MCP subagent（sentry-errors）是更高一层的 agent，按顺序调用多个 MCP server 来完成一个工作流 —— 收告警 → 拉 issue → 拉日志 → 拉 dashboard 面板 → 写三角化笔记。**server 是原语；subagent 是组合好的工作流**。通常两个都跑：server 当基础设施，subagent 当自动化。",{"q":114,"a":115},"这个 pack 能配合我现有的 on-call 工具（PagerDuty \u002F Opsgenie \u002F 自建）吗？","可以，加一个 shim。sentry-errors subagent 和 Grafana OnCall 集成都假设有一个告警源 —— 告警一旦到了 agent（webhook、Sentry 规则、Alertmanager receiver），MCP fan-out 对告警源是无感的。PagerDuty 和 Opsgenie 都有 webhook，路由到同一个 agent endpoint 就行。**本 pack 不替代你的 paging 工具，而是在人被叫起之前先做一轮三角化**。",{"@context":117,"@type":118,"name":13,"description":119,"numberOfItems":120,"inLanguage":121},"https:\u002F\u002Fschema.org","ItemList","9 个 MCP server 和 subagent 按推荐安装顺序：日志查询 MCP（SigNoz \u002F Axiom）、dashboard MCP（Grafana \u002F Datadog）、错误 MCP（Sentry + 自动三角化 subagent）、CLI 桥（Pup）、LLM trace MCP（langfuse-mcp \u002F Monoscope）—— 你监控栈的 agent 接入层。",9,"zh-CN",[123,127,131],{"url":124,"anchor":125,"reason":126},"\u002Fzh\u002Fai-tools-for\u002Fobservability","可观测性 AI 工具集","本 pack 的 MCP 就是这个目录的 agent 接入入口",{"url":128,"anchor":129,"reason":130},"\u002Fzh\u002Ftopics","浏览其他主题 pack","和 log-analysis-search、deploy-monitor-observability pack 组合成完整栈",{"url":132,"anchor":133,"reason":134},"\u002Fzh\u002Ffeatured","TokRepo 精选资产","这 9 个 MCP 属于更大的精选目录",[136,140,144],{"claim":137,"source_name":138,"source_url":139},"Model Context Protocol 是连接工具和 LLM agent 的开放标准","Model Context Protocol 规范","https:\u002F\u002Fmodelcontextprotocol.io\u002F",{"claim":141,"source_name":142,"source_url":143},"Grafana 提供第一方 MCP server，暴露 dashboard、告警、OnCall 工具","Grafana 官方文档","https:\u002F\u002Fgrafana.com\u002Fdocs\u002Fgrafana\u002Flatest\u002F",{"claim":145,"source_name":146,"source_url":147},"Sentry 提供官方 MCP server，让 AI agent 查询错误和 issue","Sentry MCP server 文档","https:\u002F\u002Fdocs.sentry.io\u002F",920,"2026-05-22T00:00:00Z"]