mcp-debug-claude-desktop
An instrumented MCP server for debugging lost tool calls in Claude Desktop, using ping, echo, sized_echo, and slow_echo tools to record and analyze the full protocol flow.
README
mcp-debug-claude-desktop
Filed as anthropics/claude-code#83908.
A minimal instrumented MCP server for answering one question: when a tool call goes missing in Claude Desktop, where does it go?
Claude Desktop reports lost tool calls as "the local MCP server may be unresponsive, crashed, or not running." This repository checks that claim instead of believing it. For every message it records whether bytes arrived on stdin, whether they parsed, whether a handler ran, whether a response was written, and whether that write went through.
What it found
A tool call is dispatched only if its chat is the active chat at the moment of dispatch — which happens 5–45 seconds after you submit. Open a different chat in that window and the call is discarded: nothing reaches the server, no retry, no log entry anywhere in the client. Four minutes later the user is told the local server has crashed.
Ruled out with the measurement that refuted each: server crash, back-pressure, the approval dialog, deferred tool loading, Electron background throttling. Details in BUG_REPORT.md.
Quick start
npm install
npm run build
npm run test:backpressure # verify the instrument before trusting it
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"mcp-debug": {
"command": "/absolute/path/to/node",
"args": ["/absolute/path/to/mcp-debug-claude-desktop/dist/server.js"]
}
}
}
Both paths absolute; command is the output of which node. Restart Claude Desktop, use a tool,
then npm run analyse:
rpc_id method / tool stdin parsed handler stdout verdict
0 initialize 245 B yes — 162 B ok complete
2 tools/call echo 108 B yes 0 ms 87 B ok complete
pid ppid requests responses blocked role
37265 37257 10 10 0 active — 8 tool call(s)
37270 37266 2 2 0 handshake only — no tool call ever
4 of 12 expected tool calls never reached any server process.
PROTOCOL.md is the full reproduction procedure.
The tools
| Tool | Purpose |
|---|---|
ping |
Smallest round trip. Returns a per-process sequence number — an unbroken sequence proves the server never restarted |
echo |
Returns its text argument, so each chat is identifiable in the capture |
sized_echo |
Returns size_kb kilobytes of filler |
slow_echo |
Returns after delay_ms, separating a slow handler from a lost message |
What gets recorded
One JSON object per line in captures/. Never to stdout — stdout is the protocol channel.
| Record | Contents |
|---|---|
session |
pid, ppid, argv, node version, start time |
raw |
Bytes in both directions, before any parsing |
parsed |
The JSON-RPC message those bytes became |
unparseable |
Bytes that never became a message |
handler |
Start and end of each tool handler |
write |
The synchronous result of process.stdout.write(), with writableNeedDrain |
blocked |
Emitted immediately when a write does not go through |
flushed |
When the write completed, and how long it waited |
lifecycle |
stdin end/close/error, stdout EPIPE, signals |
Two details the measurement depends on. stdin is teed through a PassThrough, with the capture
listener registered before the transport's, so no byte is consumed before it is recorded.
process.stdout.write is replaced, not wrapped — a wrapper's own buffer would report success
while the real stdout was blocked, and the blocked record is written synchronously, because a
client that stopped reading never fires drain.
Verifying the instrument
npm run test:backpressure
Runs the server against a consumer that never reads, and asserts the detector fires: 24 tool calls, 24 blocked writes, 1 delivered. An instrument that has never triggered cannot be trusted when it stays silent.
Layout
src/server.ts the MCP server and its four tools
src/capture.ts the capture file and the record formats
src/wire.ts stdin tee and the stdout replacement
src/analyse.ts npm run analyse
src/test/backpressure.ts npm run test:backpressure
samples/ a small capture, plus the reference run behind the report
issues/ the reports as submitted
Privacy
The tools are synthetic: they return fixed strings, filler, or their own arguments. captures/ is
gitignored; the committed samples have paths replaced. Claude Desktop's own logs are different —
they contain organisation and conversation UUIDs, and claude_desktop_config.json may contain API
keys. See PROTOCOL.md before attaching anything.
Requirements
Node ≥ 20. One dependency, @modelcontextprotocol/sdk, pinned to 1.30.0.
MIT — see LICENSE.
I am currently looking for a new position; if this is the kind of work you need, you can reach me at sascha.rose@gmail.com.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。