MCP stdio工具报错：mcp.shared.exceptions.McpError: Connection closed

[sfsflm]

求助一个问题，通过MCP stdio的方式引入了一个工具，今天在使用的时候出现了mcp.shared.exceptions.McpError: Connection closed的错误。日志中有“error: Failed to fetch: https://pypi.org/simple/tabulate/”，不知道是否是这个影响的？ 这也是第一次遇到这个问题，之前没遇到过这个问题。

配置是：

    "semantic-scholar": {
      "enabled": true,
      "type": "stdio",
      "command": "uvx",
      "args": [
        "semantic-scholar-mcp"
      ],
      "env": {
        "SEMANTIC_SCHOLAR_API_KEY": "$SEMANTIC_SCHOLAR_API_KEY"
      },
      "description": "Academic paper search tool"
    },

2026-06-12 02:42:37 - mcp.client.streamable_http - INFO - Received session ID: 8c3b1a8263c84018a3ff8829e06467b6
2026-06-12 02:42:37 - mcp.client.streamable_http - INFO - Negotiated protocol version: 2025-11-25
Downloading cryptography (4.5MiB)
Downloading pygments (1.2MiB)
Downloading networkx (2.0MiB)
Downloading pydantic-core (2.0MiB)
Downloading pymupdf (23.8MiB)
Downloading numpy (15.9MiB)
Downloading pymupdf-layout (15.1MiB)
Downloading onnxruntime (17.3MiB)
 Downloading pygments
 Downloading networkx
 Downloading pydantic-core
 Downloading cryptography
 Downloading pymupdf-layout
 Downloading numpy
 Downloading onnxruntime
 Downloading pymupdf
Installed 55 packages in 46ms
{"timestamp": "2026-06-12T02:50:26.945554+00:00", "level": "INFO", "logger": "semantic_scholar_mcp.instruction_loader", "message": "Loaded all tool instructions", "module": "logging", "function": "_log", "line": 164, "taskName": null, "total_tools": 24}
 中间省略日志
INFO:     10.233.127.90:42962 - "GET /api/v1/auth/setup-status HTTP/1.1" 200 OK
INFO:     10.233.127.90:42962 - "GET /api/v1/auth/setup-status HTTP/1.1" 200 OK
INFO:     10.233.127.90:41054 - "GET /api/v1/auth/setup-status HTTP/1.1" 200 OK
INFO:     10.233.127.90:41054 - "GET /api/v1/auth/setup-status HTTP/1.1" 200 OK
INFO:     10.233.127.90:41128 - "GET /api/v1/auth/setup-status HTTP/1.1" 200 OK
INFO:     10.233.127.90:41128 - "GET /api/v1/auth/setup-status HTTP/1.1" 200 OK
INFO:     10.233.127.90:35196 - "GET /api/v1/auth/setup-status HTTP/1.1" 200 OK
INFO:     10.233.127.90:35196 - "GET /api/v1/auth/setup-status HTTP/1.1" 200 OK
error: Failed to fetch: `https://pypi.org/simple/tabulate/`
  Caused by: Request failed after 3 retries
  Caused by: error sending request for url (https://pypi.org/simple/tabulate/)
  Caused by: operation timed out
2026-06-12 03:36:02 - deerflow.agents.middlewares.tool_error_handling_middleware - ERROR - Tool execution failed (async): name=semantic-scholar_search_papers id=tool-ccab998078f04b21a76c30cfacc56cd6
Traceback (most recent call last):
  File "/app/backend/packages/harness/deerflow/agents/middlewares/tool_error_handling_middleware.py", line 116, in awrap_tool_call
    result = await handler(request)
             ^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain/agents/factory.py", line 669, in call_inner
    return await inner(req, execute)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/packages/harness/deerflow/agents/middlewares/clarification_middleware.py", line 216, in awrap_tool_call
    return await handler(request)
           ^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1196, in execute
    return await self._execute_tool_async(req, input_type, config)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1151, in _execute_tool_async
    content = _handle_tool_error(e, flag=self._handle_tool_errors)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 432, in _handle_tool_error
    content = flag(e)  # type: ignore [assignment, call-arg]
              ^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 389, in _default_handle_tool_errors
    raise e
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1103, in _execute_tool_async
    response = await tool.ainvoke(call_args, config)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/structured.py", line 70, in ainvoke
    return await super().ainvoke(input, config, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/base.py", line 652, in ainvoke
    return await self.arun(tool_input, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/base.py", line 1131, in arun
    raise error_to_raise
  File "/app/backend/packages/harness/deerflow/agents/middlewares/clarification_middleware.py", line 216, in awrap_tool_call
    return await handler(request)
           ^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1196, in execute
    return await self._execute_tool_async(req, input_type, config)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1151, in _execute_tool_async
    content = _handle_tool_error(e, flag=self._handle_tool_errors)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 432, in _handle_tool_error
    content = flag(e)  # type: ignore [assignment, call-arg]
              ^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 389, in _default_handle_tool_errors
    raise e
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1103, in _execute_tool_async
    response = await tool.ainvoke(call_args, config)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/structured.py", line 70, in ainvoke
    return await super().ainvoke(input, config, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/base.py", line 652, in ainvoke
    return await self.arun(tool_input, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/base.py", line 1131, in arun
    raise error_to_raise
  File "/app/backend/packages/harness/deerflow/agents/middlewares/clarification_middleware.py", line 216, in awrap_tool_call
    return await handler(request)
           ^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1196, in execute
    return await self._execute_tool_async(req, input_type, config)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1151, in _execute_tool_async
    content = _handle_tool_error(e, flag=self._handle_tool_errors)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 432, in _handle_tool_error
    content = flag(e)  # type: ignore [assignment, call-arg]
              ^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 389, in _default_handle_tool_errors
    raise e
  File "/app/backend/.venv/lib/python3.12/site-packages/langgraph/prebuilt/tool_node.py", line 1103, in _execute_tool_async
    response = await tool.ainvoke(call_args, config)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/structured.py", line 70, in ainvoke
    return await super().ainvoke(input, config, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/base.py", line 652, in ainvoke
    return await self.arun(tool_input, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/base.py", line 1131, in arun
    raise error_to_raise
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/base.py", line 1097, in arun
    response = await coro_with_context(coro, context)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/structured.py", line 124, in _arun
    return await self.coroutine(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/packages/harness/deerflow/mcp/tools.py", line 135, in call_with_persistent_session
    session = await pool.get_session(server_name, thread_id, connection)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/base.py", line 1097, in arun
    response = await coro_with_context(coro, context)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain_core/tools/structured.py", line 124, in _arun
    return await self.coroutine(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/packages/harness/deerflow/mcp/tools.py", line 135, in call_with_persistent_session
    session = await pool.get_session(server_name, thread_id, connection)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/packages/harness/deerflow/mcp/session_pool.py", line 213, in get_session
    return await asyncio.shield(join)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/packages/harness/deerflow/mcp/session_pool.py", line 113, in _run_session
    await session.initialize()
  File "/app/backend/.venv/lib/python3.12/site-packages/mcp/client/session.py", line 171, in initialize
    result = await self.send_request(
             ^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/mcp/shared/session.py", line 306, in send_request
    raise McpError(response_or_error.error)
mcp.shared.exceptions.McpError: Connection closed
2026-06-12 03:36:02 - deerflow.agents.middlewares.tool_error_handling_middleware - ERROR - Tool execution failed (async): name=semantic-scholar_search_papers id=tool-16b2a50232904de397b3b648b350b1af
Traceback (most recent call last):
  File "/app/backend/packages/harness/deerflow/agents/middlewares/tool_error_handling_middleware.py", line 116, in awrap_tool_call
    result = await handler(request)
             ^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain/agents/factory.py", line 669, in call_inner
    return await inner(req, execute)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/packages/harness/deerflow/agents/middlewares/tool_error_handling_middleware.py", line 116, in awrap_tool_call
    result = await handler(request)
             ^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/.venv/lib/python3.12/site-packages/langchain/agents/factory.py", line 669, in call_inner
    return await inner(req, execute)
           ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/app/backend/packages/harness/deerflow/agents/middlewares/clarification_middleware.py", line 216, in awrap_tool_call
    return await handler(request)
           ^^^^^^^^^^^^^^^^^^^^^^

[fancyboi999]

感觉是网络问题哈，拉包网络问题。 看日志就是 uvx semantic-scholar-mcp 启动时访问 PyPI 拉 tabulate 超时，导致 MCP stdio 子进程退出，所以 deer-flow 侧才报 Connection closed；建议在同容器里测 curl https://pypi.org/simple/tabulate/，或给 uv 配镜像/预装依赖，别让线上每次靠 uvx 临时拉包。

[P4ST4S]

看日志的话，主要问题大概率不是 deer-flow 的 MCP 协议逻辑本身，而是 stdio 子进程在启动或重新初始化时失败退出了。

关键线索是这里：

error: Failed to fetch: `https://pypi.org/simple/tabulate/`
Caused by: Request failed after 3 retries
Caused by: operation timed out
...
mcp.shared.exceptions.McpError: Connection closed

uvx semantic-scholar-mcp 启动时需要解析和安装 Python 包。如果它在拉取依赖时访问 PyPI 超时，子进程会退出。对 MCP client 来说，stdio 连接就被关闭了，所以 deer-flow 这边看到的是 Connection closed。

我会先在运行 deer-flow 的同一个容器或同一台机器里验证这几件事：

curl -I https://pypi.org/simple/tabulate/
uvx semantic-scholar-mcp --help

如果第二个命令也会卡住或失败，那就基本可以确认是运行时拉包失败导致 MCP server 没有稳定启动。

建议不要在线上每次 tool call 或 session 初始化时依赖 uvx 临时下载依赖。更稳的做法是把 MCP server 预装好：

uv tool install semantic-scholar-mcp

然后配置里直接运行安装后的命令，而不是每次用 uvx 临时解析：

{
  "semantic-scholar": {
    "enabled": true,
    "type": "stdio",
    "command": "semantic-scholar-mcp",
    "args": [],
    "env": {
      "SEMANTIC_SCHOLAR_API_KEY": "$SEMANTIC_SCHOLAR_API_KEY"
    },
    "description": "Academic paper search tool"
  }
}

如果是在 Docker 里跑，最好在镜像构建阶段完成安装，避免运行时访问 PyPI：

RUN uv tool install semantic-scholar-mcp

如果你的环境访问 PyPI 不稳定，可以给 uv 配置可访问的 index mirror。按你们使用的 uv 版本，可以用类似：

UV_DEFAULT_INDEX=https://pypi.tuna.tsinghua.edu.cn/simple

或者在启动命令里显式指定 index。

另外还有一个 stdio MCP 常见坑也建议顺手检查：MCP server 的普通日志不能写到 stdout。stdout 必须只输出 JSON-RPC 消息，日志应该写 stderr。否则即使进程没有退出，也可能污染 MCP 协议流。不过从你贴的日志看，最直接的失败点还是 PyPI fetch timeout。

所以排查顺序我建议是：

1. 在同容器内确认能访问 https://pypi.org/simple/tabulate/
2. 手动运行 uvx semantic-scholar-mcp --help
3. 改成预安装 semantic-scholar-mcp
4. 给 uv 配置稳定的 package index
5. 确认 MCP server 日志不要写 stdout

如果这样处理后 semantic-scholar-mcp 能稳定启动，Connection closed 应该也会消失。

Did this resolve it? Feel free to mark as answer if so.

Note: I wrote the technical answer and translated it with GPT-5.5.