MCP server monitoring
AI agents fail quietly when an MCP server is down or a tool changes under them. Check the server the way an agent connects to it.
What each check does
- Connects to the server URL, initializes an MCP session and lists its tools.
- Supports servers behind OAuth or a token in a header.
- Optionally calls one tool with arguments you set; discovery alone never calls tools.
When it alerts
- The server cannot be reached, the session does not initialize or listing tools fails.
- A separate alert when a tool is removed or its input or output schema changes.
Example
https://mcp.example.com/mcpGood to know
- Set the timeout between 1 and 30 seconds.
- Preview the tools before saving, to check the connection settings.
- A tool call runs in every selected region, so choose one that is safe to repeat.
On every monitor
- Alerts by email, Slack, Discord, Teams, Telegram, PagerDuty, webhooks, Lark, ntfy and the mobile app.
- Recovery thresholds, so an incident closes only after the service has answered several times in a row.
- Dependencies: a monitor that depends on another stays quiet while that one is down.
- Maintenance windows, incident history and public status pages built from the same monitors.
Try it on the Free plan
The Free plan includes 3 monitors checked every 5 minutes, with every alert channel. No card needed.
Questions
Why monitor tool schemas?
An agent's prompts and code are written against the tools as they were. A renamed argument or a removed tool breaks them without any error on the server side.
Which MCP servers can be monitored?
Remote servers reachable over HTTP. A local server that runs over stdio has no URL to check.
MCP 服务器监控
MCP 服务器宕机或工具悄悄变化时，AI Agent 会无声地失败。按照 Agent 连接它的方式检查服务器。
每次检查做什么
- 连接服务器 URL，初始化 MCP 会话并列出工具。
- 支持需要 OAuth 或请求头令牌的服务器。
- 可以用你设定的参数调用一个工具；仅做发现时绝不会调用工具。
什么时候告警
- 服务器无法访问、会话无法初始化，或列出工具失败。
- 工具被删除，或输入、输出结构发生变化时，单独发出提醒。
示例
https://mcp.example.com/mcp使用说明
- 超时可设为 1 到 30 秒。
- 保存前可以先预览工具列表，确认连接设置正确。
- 工具调用会在每个选中的区域执行，所以请选择可以重复调用的工具。
所有监控通用
- 通过邮件、Slack、Discord、Teams、Telegram、PagerDuty、Webhook、飞书、ntfy 和手机 App 告警。
- 可设置恢复阈值，服务连续多次正常后故障才结束。
- 依赖关系：依赖另一个监控的监控，在被依赖方宕机期间不会告警。
- 维护窗口、故障历史，以及由同一组监控生成的公开状态页。
常见问题
为什么要监控工具的参数结构？
Agent 的提示词和代码是按照工具原来的样子写的。参数改了名或工具被删掉，会让它们出错，而服务器那边不会有任何报错。
能监控哪些 MCP 服务器？
可以通过 HTTP 访问的远程服务器。通过 stdio 运行的本地服务器没有可检查的 URL。