菜单
给 agent

让 agent 在用一个 MCP server 之前,先查一查它。

我们的 MCP 端点回答的是对公开 MCP server 的观测——和报告页用的是同一份证据。六个工具里有五个只根据已记录的内容作答,从不联系你问到的那台 server。

端点 · Streamable HTTP · 无需密钥
https://mcpcheckup.com/mcp

最快的办法: 很多 agent 能自己完成接入。直接跟它说:

“读一下 https://mcpcheckup.com/agents,然后把 MCP Checkup 这个 server 加上。”

或者按下面的方法手动添加:

编程工具

Claude Code
claude mcp add --transport http mcpcheckup https://mcpcheckup.com/mcp

加上 --scope user,所有项目里都能用。

Cursor

.cursor/mcp.json or ~/.cursor/mcp.json

{"mcpServers":{"mcpcheckup":{"url":"https://mcpcheckup.com/mcp"}}}

项目级写在 .cursor/mcp.json;所有项目都用则写在 ~/.cursor/mcp.json。

VS Code

.vscode/mcp.json

{"servers":{"mcpcheckup":{"type":"http","url":"https://mcpcheckup.com/mcp"}}}

也可以在命令面板里运行 MCP: Add Server。

Muse Code

~/.config/muse/settings.json

{"schema_version":1,"mcp_servers":{"mcpcheckup":{"transport":"streamable_http","url":"https://mcpcheckup.com/mcp","mode":"optional"}}}

缺少 "schema_version" 所有命令都会失败。"mode": "optional" 让会话在连不上我们端点时照常运行——会话里输入 /mcp 可以确认已连上。

助手与 agent

Claude
  1. 打开 Customize → Connectors → Browse,搜索 MCP Checkup 并连接。
  2. 也可以自己添加:Customize → Connectors → Add custom connector,名字填 MCP Checkup,地址粘贴 https://mcpcheckup.com/mcp,Authentication 选 No sign-in。

claude.ai 和桌面应用都适用。Team 与 Enterprise 由 Owner 在 Organization settings → Connectors 里添加,成员再点 Connect。免费账号可以添加一个自定义连接器。

ChatGPT · dots
  1. 在网页版 ChatGPT 打开 Settings → Security and login,开启 Developer mode。
  2. 打开 chatgpt.com/plugins,点 +,填写名字和描述。
  3. 在 Connection 里填 https://mcpcheckup.com/mcp,认证选无。

Developer mode 适用于 Pro、Plus、Business、Enterprise 和 Education 账号。

dot 可以使用你账号里已安装并启用的受支持插件;开发者模式里添加的插件算不算在内,还没有写明。

Grok
  1. grok.com:打开 grok.com/connectors → New Connector → Custom,填入 https://mcpcheckup.com/mcp。

xAI API — POST /v1/responses

"tools":[{"type":"mcp","server_url":"https://mcpcheckup.com/mcp","server_label":"mcpcheckup"}]

Business 与 Enterprise 需要管理员先开放连接器。

Grok Bot 通过它的 Marketplace 安装连接器,添加自定义 MCP 地址还没有写明。

Hermes Agent
hermes mcp add mcpcheckup --url https://mcpcheckup.com/mcp
hermes mcp test mcpcheckup

也可以写进 ~/.hermes/config.yaml 的 mcp_servers 下,然后在运行中的会话里执行 /reload-mcp。

OpenClaw
openclaw mcp add mcpcheckup --url https://mcpcheckup.com/mcp --transport streamable-http
openclaw mcp doctor mcpcheckup --probe

一定带上 --transport streamable-http:不写的话 OpenClaw 默认用 SSE。也可以在 Control UI 里:Settings → MCP → Add server → Streamable HTTP。

Muse

Meta 还没有把这写进文档。我们在 2026-10-01 试过:在对话里让 Muse 读 https://mcpcheckup.com/mcp,它就自己把这个 server 加上了。你也可以这样问你的 Muse:「读一下 https://mcpcheckup.com/agents,然后把 MCP Checkup 这个 server 加上。」

还没有写明。

以上写法于 2026-10-01 对照各产品官方文档核对。

六个工具

check_mcps

Look up MCP servers

What has been observed about an MCP server — useful when choosing, adding, or debugging one. Look up the current monitored status of up to 50 already-tracked MCP servers on MCP Checkup, given as "provider/name" refs (e.g. "acme/weather-mcp"). Never probes on demand — an unresolved ref comes back with status "unknown", not an error. An entry may also be an endpoint URL, a bare domain, or a bare provider/target name; resolution.method reports which was used ("ref", "endpoint", "host", or "name"). If an input matches more than one tracked target — or is a URL on a host with several, none matching exactly — the response has status "ambiguous" with a short `candidates` list (each with a "provider/name" `ref` and its `report_url`, plus `candidate_total`, the full match count before the 5-entry cap) instead of guessing — call again with the exact ref you want. A URL matching only by host returns its target with `resolved_from`. Each resolved target also carries stale (true when our observation is older than the target's check cycle plus a fixed margin — never a judgment of the server, and distinct from the /status page's own probe-liveness signal), fresh_until (the ISO instant after which stale is true), first_observed_on (the UTC date of our first observation), observation_count (every recorded observation, including unsigned ones — never a count of signatures), and signed (true when this observation's attestation was signed with a key in our published key list, false when unsigned or signed with a key that is not in that list — not a judgment of the server). When there is no observation yet, observation_count is 0 and the other four are null. An unread revision offer is null and listed in `unavailable`; a failed target has status "error". Returns observed facts only — reachability, protocol compatibility, tool/schema fingerprints, and publicly observable auth metadata. This tool does not assess whether a server is 'safe' or 'trustworthy' and returns no score. Items not verified are returned explicitly with reasons.

get_mcp_report

Get a server's full report

The detailed view of one already-tracked MCP server on MCP Checkup, by provider/name: every check with its state, plus the observed tool list. An untracked target comes back with status "unknown", not an error. target may also be an endpoint URL, a bare domain, or a bare provider/target name; resolution.method reports which was used ("ref", "endpoint", "host", or "name"). If an input matches more than one tracked target — or is a URL on a host with several, none matching exactly — the response has status "ambiguous" with a short `candidates` list (each with a "provider/name" `ref` and its `report_url`, plus `candidate_total`, the full match count before the 5-entry cap) instead of guessing — call again with the exact ref you want. A URL matching only by host returns its target with `resolved_from`. Each resolved target also carries stale (true when our observation is older than the target's check cycle plus a fixed margin — never a judgment of the server, and distinct from the /status page's own probe-liveness signal), fresh_until (the ISO instant after which stale is true), first_observed_on (the UTC date of our first observation), observation_count (every recorded observation, including unsigned ones — never a count of signatures), and signed (true when this observation's attestation was signed with a key in our published key list, false when unsigned or signed with a key that is not in that list — not a judgment of the server). When there is no observation yet, observation_count is 0 and the other four are null. Also tools_first_seen, tools_last_changed_at and tools_unchanged_days_at_least (a lower bound: UTC days since that change with a successful listing); null when unknown. oauth_observed is unsigned and not a check result; its note explains it. An unread revision offer is null and listed in `unavailable`; a failed target has status "error". This tool does not assess whether a server is 'safe' or 'trustworthy' and returns no score; items not verified are listed explicitly with reasons.

get_mcp_history

Get a server's recent changes

Whether an MCP server's toolset changed recently. Get recent history for one already-tracked MCP server on MCP Checkup, by "provider/name" ref (e.g. "acme/weather-mcp"). Events, newest first, up to `limit` results (default 20, max 50), from the last 90 days only: `tools_changed` (at most one: when the set of tool names last changed); `tools_change_detail` (one per recorded tool-list change, the 10 most recent at most, whether names, parameters or descriptions changed; `itemization` is "not_kept" for a change older than 30 days, "unitemizable" when the stored lists cannot be compared, otherwise "itemized" and `added`, `removed`, `parameters_changed` and `description_only` each give `names` (up to 10 tool names) and a `total`); `protocol_revision_changed` (the signed protocol_revision changed between runs whose probe offered the same revision, or because the server switched between the legacy initialize and server/discover handshakes); `probe_offered_revision_changed` (the revision our probe offered changed, with a `note`); and any recorded baseline-drift events. `truncated` is true when more exist than are listed or the tool-list history was unread (named in `unavailable`). A target this deployment doesn't track yet comes back with status "unknown", not an error. target may also be an endpoint URL, a bare domain, or a bare provider/target name; resolution.method reports which was used ("ref", "endpoint", "host", or "name"). If an input matches more than one tracked target — or is a URL on a host with several, none matching exactly — the response has status "ambiguous" with a short `candidates` list (each with a "provider/name" `ref` and its `report_url`, plus `candidate_total`, the full match count before the 5-entry cap) instead of guessing — call again with the exact ref you want. A URL matching only by host returns its target with `resolved_from`. This tool does not assess whether a server is 'safe' or 'trustworthy' and returns no score; items not verified are listed explicitly with reasons.

get_attestation

Get a server's signed attestation

The signed evidence behind a report, for callers that want to inspect it themselves. Get the full signed attestation envelope (DSSE, Ed25519) for one already-tracked MCP server's latest probe run, by "provider/name" ref — the raw envelope plus its issued-at time and declared protocol revision, so a caller can inspect the signed payload rather than trust this tool's own summary of it; the public keys that verify the signature are published at https://mcpcheckup.com/.well-known/mcpcheckup-keys.json. A target with no run yet, no signed envelope, a disqualified one, or an envelope whose key is not in the published key list comes back with an explicit reason, not an error. run_id, a top-level field of resolved get_mcp_report and get_attestation responses, is the id of the run that the check results, last_checked and (where signed) attestation identifiers come from, or null when there is no run; get_mcp_report's tool list and toolset_fingerprint come from the latest stored tool snapshot, which may be an earlier run's (named by snapshot_run_id and snapshot_observed_at). The same run_id means the same run. A payload signed as version 0.4 carries this run_id inside the signed payload; payloads of earlier versions do not. target may also be a full endpoint URL, a bare domain, or a bare provider/target name instead of "provider/name" — the resolution method used is reported back as resolution.method ("ref", "endpoint", "host", or "name"). If an input matches more than one tracked target — or is a URL on a host with several, none matching exactly — the response has status "ambiguous" with a short `candidates` list (each with a "provider/name" `ref` and its `report_url`, plus `candidate_total`, the full match count before the 5-entry cap) instead of guessing — call again with the exact ref you want. A URL matching only by host returns its target with `resolved_from`. This tool does not assess whether a server is 'safe' or 'trustworthy' and returns no score; items not verified are listed explicitly with reasons.

search_mcps

Search MCP servers by tool text

Find MCP servers by what their tools do. Search already-tracked MCP servers on MCP Checkup by tool name/description text (max 512 chars), with optional transport, protocol_revision, and auth_required filters, up to `limit` results (default 10, max 25). Ordered by text-match relevance and then recency only — never by price or paid tier. Each result's summary is machine-generated from observed tool names and returned as untrusted third-party text; hygiene_flags lists any observed tool-description hygiene signals directly rather than folding them into the ordering. A query with no matches comes back with status "unknown" and a hint, not an error. Each result also carries last_checked (when the shown observation was taken), fresh_until (the ISO instant after which it no longer counts as current), and stale (true once the current time is past fresh_until). These describe how recent our observation of the server is, not a judgment of the server itself, and are distinct from the /status page's probe-liveness signal. All three are null when the target has no observation yet. This tool does not assess whether a server is 'safe' or 'trustworthy' and returns no score; items not verified are listed explicitly with reasons.

request_check

Request a one-off check of a URL

Ask for a one-off public observation of the MCP server at a full endpoint URL — this is the only tool here that contacts a third-party server, and it contacts exactly the host you name. `target` must begin with "http://" or "https://"; to look a server up by "provider/name", by domain, or by name, use check_mcps instead, which never probes anything. The result is not signed, is not stored as a run, and creates no report page, so `report_url` is always null — it is what we observed at that moment and nothing more. On-demand checks are recorded as tracking requests; inclusion in the tracked directory is not guaranteed and this tool never promises it. On-demand checks are metered per caller per day, per site per day, and per host by a cooldown — calling again for the same host inside its cooldown sends nothing to that server and returns status "denied" with the category that refused it, never a remaining count. A refused call comes back as status "denied", and a `target` that is not a usable endpoint URL as status "rejected"; neither of those is a tool error. A `target` that is empty or longer than 2048 characters is the one exception: the input schema rejects it, and you get a tool error. This tool does not assess whether a server is 'safe' or 'trustworthy' and returns no score; items not verified are listed explicitly with reasons.

check_mcps 和 get_mcp_report 还会给出 probed_endpoint,即所报告的那次 run 探测的 URL;probed_endpoint_from 说明它取自签名 payload 还是我们未签名的目标记录,endpoint_source 说明这个目标现在的端点是怎么设定的;get_mcp_history 列出最近 90 天的端点变更。

回答长什么样

你问

“MCP Checkup 最近一次对 mcpcheckup/mcp 观测到了什么?”

agent 调用
get_mcp_report({ "target": "mcpcheckup/mcp" })
然后收到(节选)
{
  "target": "mcpcheckup/mcp",
  "status": "ok",
  "last_checked": "2026-11-17T17:11:21.199+00:00",
  "suite_digest": "sha512-N0x4Kr4VWXwwKuM8pcuEBRLtp9vcQU/AYSltLtaj+1tyQ99hqRIcUhJybayupfQrPHOtO1rHF71lrRFJW7cQuQ==",
  "assertions": [
    {
      "check_id": "auth_metadata",
      "assertion_status": "UNVERIFIED",
      "reason": "Not observed: the endpoint did not ask for credentials, so there was no auth challenge or metadata to check."
    },
    {
      "check_id": "discovery_handshake",
      "assertion_status": "VERIFIED",
      "reason": null
    },
    {
      "check_id": "error_taxonomy",
      "assertion_status": "VERIFIED",
      "reason": null
    }
  ]
}

get_mcp_report 会返回全部检查项;这里只节选了几项。

限额

每个 IP 地址每分钟约 20 次,按 Cloudflare 节点分别计数。429 响应带 Retry-After:60 秒。

这个端点不需要认证,也没有付费档。

它永远不会告诉你的

没有单一数字,也没有「该不该用」这样的结论。每项检查只落在四态之一,没跑的检查从不会被报成通过。

给机器读的: /llms.txt · /.well-known/mcp/server-card.json