mcpspan

io.github.mcpspan由 GitHub 账号 mcpspan 发布v0.14.0更新于 Nov 6, 2026

Ask your agent about your MCP servers: failing tools, releases, clients, sessions.

已验证Streamable HTTP可网页运行开发者工具安全与监控数据与分析

概览

AI 生成的概览

让助手查询 mcpspan 中你自己 MCP 服务器的分析数据：失败的工具、错误率、客户端、版本与会话。

可以做什么
  • 列出账号上报的服务器，并查看某个服务器的调用量、失败数、错误率、响应时间与客户端构成
  • 按调用量、错误、响应时间或名称排列服务器的工具，并深入查看单个工具的失败原因、被拒参数、回答大小、定义变更与重复调用
  • 查看最近的失败调用、代理请求但服务器没有的工具，以及连接过的客户端
  • 并排比较各版本，查看代理看到却从未调用的工具，并逐个调用检查会话
  • 查看代理卡在哪里、如何恢复以及接下来尝试了什么
适用场景
当你运行接入 mcpspan 的 MCP 服务器，想用自然语言问代理某个工具为何失败、新版本与上一版相比如何、哪个客户端或工具出了问题，而不必手动翻看仪表盘时适用。
运行要求
两种方式：连接 mcpspan API 的 /mcp 远程地址并用 OAuth 登录；或用 npx 启动本地进程（需 Node 20 或更新版本），通过 MCPSPAN_READ_TOKEN 提供读取令牌。自托管安装需设置 MCPSPAN_URL，命令行方式需 mcpspan 0.9.0 或更新版本，地址方式需 0.13.0 或更新版本。还需要一个有数据的 mcpspan 账号。
安装前请注意
MCPSPAN_READ_TOKEN 中的读取令牌是只显示一次的凭据，泄露后应在设置中撤销；若代理通过互联网访问自托管安装，请在前面加 HTTPS，因为令牌会随每次请求传输。OAuth 授权按应用授予，可在已连接应用中随时断开。所有工具只读，不会修改任何内容，也不会把数据发往你的 mcpspan 之外。

安装

在 SourceWeft 中

  1. 打开 控制台中的 mcpspan，将其添加到工作区。
  2. 连接 所需的 凭据。它们会被私密存储，绝不会在公开页面上展示。
  3. 为需要使用其工具的对话启用该服务。

Web executable，通过 Streamable HTTP。 远程服务在工作区中配置后即可从网页运行时运行。

交给智能体安装

把这段提示词粘贴到 Claude Code、Codex、Cursor 等智能体，或 SourceWeft 对话中。智能体会读取该服务器的安装指南，先向你展示发布者、端点和登录方式，经你同意后再把它添加到你的客户端。密钥不会写进任何命令。

阅读 https://sourceweft.com/mcp/io.github.mcpspan%2Fmcpspan/install.md，并按其中说明把这个 MCP 服务器接入我的客户端。修改任何配置之前，先告诉我它的发布者、端点和登录方式，等我确认。不要把 API 密钥写进命令或我会提交的文件；问我密钥应从哪里获取。

查看智能体所依据的安装指南

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。 客户端会在首次连接时完成 OAuth 登录。

{
  "mcpServers": {
    "mcpspan": {
      "type": "http",
      "url": "https://api.mcpspan.com/mcp"
    }
  }
}

README

mcpspan-mcp

Ask your AI agent about your MCP servers. Claude Code, Cursor or any other MCP client answers from your mcpspan numbers:

  • Which tools failed most this week, and why?
  • How does 1.4.1 compare with 1.4.0?
  • What do agents ask for that my server does not have?
  • Does Cursor fail more often than Claude?
  • Which tools are agents shown but never call?

It sees what you see in the dashboard and changes nothing. It sends nothing anywhere but your mcpspan.

There are two ways to add it: by address, which needs nothing installed, or as a command your agent starts.

Try it on the demo

No account needed: the live demo answers any token. In Claude Code:

sh
claude mcp add --transport http mcpspan-demo https://demo-api.mcpspan.com/mcp \  --header "Authorization: Bearer demo"

Then ask about the demo's server, Flights: what went wrong in the last day?

By address

The MCP server runs inside mcpspan itself, at /mcp on its API. Add that address to your agent, and the first time it connects it opens mcpspan in your browser: sign in, check which app is asking, and allow it. Nothing to copy.

mcpspan Cloud

The address is https://api.mcpspan.com/mcp.

AgentHow
claude.ai, Claude Desktop, the Claude appsSettings → Connectors → Browse connectors, find mcpspan, and Connect. Or Add custom connector, with the address.
ChatGPTIn its connector settings (developer mode), a new connector with the address, signing in with OAuth.
Claude Codeclaude mcp add --transport http mcpspan https://api.mcpspan.com/mcp, then /mcp in Claude Code to sign in.
CursorIn mcp.json: {"mcpServers": {"mcpspan": {"url": "https://api.mcpspan.com/mcp"}}}
VS CodeIn .vscode/mcp.json: {"servers": {"mcpspan": {"type": "http", "url": "https://api.mcpspan.com/mcp"}}}
OthersWhatever they call a remote, HTTP or Streamable HTTP server, with the address.

The apps you allowed are listed in Settings → Connected apps, where one click disconnects an app and ends its access at once.

An agent that cannot sign in with OAuth can send a read token instead (see Get a token), as a header:

sh
claude mcp add --transport http mcpspan https://api.mcpspan.com/mcp \  --header "Authorization: Bearer mcpspan_read_..."

Self-hosted

The address is your API's, the one your MCP servers report to, with /mcp: http://localhost:6271/mcp on the machine itself. Agents on that machine, such as Claude Code or Cursor, connect as above.

An app that connects from elsewhere, such as claude.ai or ChatGPT, needs that address on HTTPS and reachable from the internet, and the API told where people open your dashboard, since that is where it sends them to allow it: set MCPSPAN_DASHBOARD_URL in .env, for instance https://mcpspan.example.com.

Needs mcpspan 0.13.0 or newer.

As a command

The same tools, started by your agent with npx, reading with a read token. For agents that only start commands, or to keep it on your machine.

Get a token

In mcpspan, open Settings → Read tokens and make one. It is shown once, already filled into the lines below, ready to copy.

mcpspan Cloud

Claude Code:

sh
claude mcp add mcpspan -e MCPSPAN_READ_TOKEN=mcpspan_read_... -- npx -y mcpspan-mcp

Cursor, Claude Desktop, Windsurf and others, in their MCP settings:

json
{  "mcpServers": {    "mcpspan": {      "command": "npx",      "args": ["-y", "mcpspan-mcp"],      "env": { "MCPSPAN_READ_TOKEN": "mcpspan_read_..." }    }  }}

Self-hosted

The same, with the address your MCP servers report to:

sh
claude mcp add mcpspan \  -e MCPSPAN_READ_TOKEN=mcpspan_read_... \  -e MCPSPAN_URL=http://localhost:6271 \  -- npx -y mcpspan-mcp
json
{  "mcpServers": {    "mcpspan": {      "command": "npx",      "args": ["-y", "mcpspan-mcp"],      "env": {        "MCPSPAN_READ_TOKEN": "mcpspan_read_...",        "MCPSPAN_URL": "http://localhost:6271"      }    }  }}

Needs mcpspan 0.9.0 or newer. If the agent reaches your installation over the internet, put HTTPS in front of it: the token travels with every question.

Settings

Variable
MCPSPAN_READ_TOKENThe read token. Required.
MCPSPAN_URLYour installation's API, for self-hosted mcpspan. https://api.mcpspan.com when left out.

Needs Node 20 or newer.

Tools

Every tool only reads. Each takes the server by name or id (left out when the account has one) and a window: days back from now (7 when left out), or from and to.

ToolAnswers
list_serversThe servers the account reports on.
get_overviewCalls, failures, error rate, response times and clients.
list_toolsTools ranked by calls, errors, response time or name.
get_toolOne tool: why it failed, refused arguments, answer sizes, definition changes, repeats, and the call before each failure.
list_errorsThe latest failed calls, with the kind of failure and the message.
list_unknown_toolsTools agents asked for that the server lacks, the closest one it has, and what the agent did next.
list_clientsEach client's calls, error rate and response times.
compare_versionsThe server's releases side by side.
get_tool_listingWhat agents were shown: tools never used, and what their definitions cost.
list_sessionsRecent sessions, with failures and repeated calls.
get_sessionOne session's calls, in order.
get_where_agents_get_lostWhat came of each session, the tools agents got stuck at and what they did next; for one tool, by client.

A token can be revoked at any time under Settings → Read tokens, and an app disconnected under Settings → Connected apps; either loses access at once.

Licence

MIT

来源：packages/mcpspan-mcp/README.md，提交 4b5213f

工具

12

list_servers

List servers

The MCP servers this mcpspan account reports on, with their names and ids.

无需输入参数。

get_overview

Overview

Calls, failures, error rate, response times (median and 95th percentile) and the share of each client, for one server over a window.

输入参数todaysfromclientserver

list_tools

List tools

A server's tools ranked by calls, errors, response time or name, each with its calls, errors, error rate and response times.

输入参数todaysfromsortlimitclientserver

get_tool

Tool details

Everything recorded about one tool: why its calls failed and the commonest error messages, which arguments the server refused, its parameters, how large its answers are, when its definition changed, how often agents repeated a call, and which call came right before the ones that repeated or failed.

输入参数todaysfromtool必填server

list_errors

List failed calls

The most recent failed calls, newest first: which tool, which client, when, what kind of failure (a handler that threw, a tool that answered with an error, refused arguments, an unknown tool) and the error message when the server records them.

输入参数todaysfromtoollimitclientserver

list_unknown_tools

List unknown tools

Tools that agents asked for but the server does not have: how often, which clients, the closest tool the server does have, and what the agent did next.

输入参数todaysfromserver

list_clients

List clients

The clients that called the server (Claude, Claude Code, Cursor, VS Code, ChatGPT and others), each with its calls, error rate and response times.

输入参数todaysfromserver

compare_versions

Compare versions

The server's versions in the window, newest first, each with when it was first seen, its calls, error rate and response times: how a release compares with the one before.

输入参数todaysfromclientserver

get_tool_listing

What agents are shown

Every tool as agents were shown it when they listed the tools: in how many sessions, how many of those called it, and how large its definition is. Finds tools agents see but never use, and definitions that cost context.

输入参数todaysfromclientserver

list_sessions

List sessions

Recent sessions (one connection each), newest first, with their calls, failures, tools used and calls an agent repeated with the same arguments.

输入参数todaysfromtoolwithlimitclientserver

get_where_agents_get_lost

Where agents get lost

Where agents got stuck and what they did about it. A time stuck is a failed tool call and the retries of the same tool straight after it; each ends recovered (the tool then worked, with meanRetries retries on average), wentElsewhere (the agent called something else, insteadMostOften), endedOnIt (the session stopped there, nothing for 30 minutes) or undecided (too recent to tell); looped counts those where a retry repeated the same arguments. Without a tool: every session by the worst that happened in it (clean, recovered, wentElsewhere, looped, endedOnFailure, undecided) and the tools agents got stuck at, worst first. With a tool: that tool's times stuck, as a whole and by client.

输入参数todaysfromtoolclientserver

get_session

Session

One session's calls in the order the agent made them, with each call's outcome and error.

输入参数todaysfromserversession必填

版本历史

4
  1. v0.14.0最新Nov 6, 2026
  2. v0.13.0Oct 29, 2026
  3. v0.12.0Oct 26, 2026
  4. v0.11.0Oct 25, 2026

评分

已满足适用检查项的 100%

该评分衡量此收录条目的完整程度，依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价，也不表示服务器安全或当前可用：可连通性是带日期的检查结果，评分也不包含安装量或用户评价。

  • 近 30 天内可连通必需25 分
  • 已声明认证方式必需10 分
  • 已列出工具15 分
  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分