stdio server ignores the first two Ctrl+C presses at a terminal (blocked stdin worker thread)

[pwesselius]

Description

Running an MCPServer over stdio from a terminal, Ctrl+C does nothing the first time, nothing visible the second time, and the third ends the process with a traceback out of threading._shutdown. This is related to #2663 but is not the same problem, and the fix proposed there (catching KeyboardInterrupt around anyio.run, #2745) does not address it: on a TTY the first Ctrl+C never raises KeyboardInterrupt.

A client that closes stdin never sees this, because EOF ends the read. It only affects people running a server by hand, which is what you do when testing one.

Reproduction

server.py:

from mcp.server.mcpserver import MCPServer

MCPServer("repro").run("stdio")

Run python server.py in a terminal and press Ctrl+C three times. Or, scripted, with a pty as stdin:

import pty, signal, subprocess, sys, time

master, slave = pty.openpty()
proc = subprocess.Popen([sys.executable, "server.py"], stdin=slave, stderr=subprocess.PIPE)
time.sleep(2)  # let it reach the stdin read
for n in range(1, 6):
    proc.send_signal(signal.SIGINT)
    try:
        proc.wait(1)
        break
    except subprocess.TimeoutExpired:
        print(f"still running 1s after SIGINT #{n}")
print(f"exit code {proc.returncode} after {n} SIGINT(s)")
print(proc.stderr.read().decode()[-400:])

Output:

still running 1s after SIGINT #1
still running 1s after SIGINT #2
exit code -2 after 3 SIGINT(s)
...
Exception ignored while joining a thread in _thread._shutdown():
Traceback (most recent call last):
  File "/usr/lib64/python3.14/threading.py", line 1583, in _shutdown
    _thread_shutdown()
KeyboardInterrupt:

Cause

stdio_server() reads stdin through anyio.wrap_file, so each readline() runs in an AnyIO worker thread. A thread blocked in readline() on a terminal cannot be cancelled.

1. First SIGINT: asyncio's handler cancels the main task. The task cannot finish cancelling, because stdin_reader is awaiting the worker thread. A faulthandler dump taken at this point shows the main thread still in run_forever → select, and the AnyIO worker thread inside the blocking call.
2. Second SIGINT: asyncio raises KeyboardInterrupt out of the loop. Interpreter shutdown then joins the same worker thread, which is still blocked.
3. Third SIGINT: interrupts that join, producing the _thread._shutdown traceback.

The comment in _claim_fd ("a worker thread can still block on this descriptor after the transport exits") suggests this is known for the descriptor's lifetime; this is the same thread seen from the user's side.

Possible fixes

• In MCPServer.run("stdio"), install a SIGINT handler for the duration of the run that ends the process without waiting for the reader thread. We use os._exit(130) in our server, which is only appropriate if nothing needs unwinding, so probably not as a library default.
• Read stdin without a thread where the platform allows it (on POSIX, wait for fd readability on the event loop and read non-blocking), so cancellation can actually complete.
• Run the blocking read with abandon_on_cancel=True in a daemon thread, so neither task cancellation nor interpreter shutdown waits on it.

Environment

MCP Python SDK 2.2.0, anyio 4.15.1, Python 3.14.7, Linux (Fedora 44). Not tested on v1.x, macOS or Windows.