`stdio_server` uses unbuffered memory streams which can cause server to block and become unresponsive

[peteski22]

Initial Checks

• I confirm that I'm using the latest version of MCP Python SDK
• I confirm that I searched for my issue in https://github.com/modelcontextprotocol/python-sdk/issues before opening this issue

Description

The MCP Python SDK's stdio_server becomes unresponsive during slow message processing operations, causing ping and/or new requests to timeout.

Observed Behaviour

• Server appears to 'freeze' and becomes unresponsive after running for extended periods
• Ping requests timeout during slow operations
• New requests cannot be processed while the server is handling long-running operations
• After 67 minutes of operation with requests every 10 seconds, the server became completely unresponsive

Expected Behaviour

• Server should remain responsive to new requests (like pings) even while processing slow operations
• Ping requests should not timeout due to message processing delays
• The server should handle concurrent requests without blocking

Suspected Root Cause

The stdio_server uses max_buffer_size=0 (supplied to anyio.create_memory_object_stream) by default, which creates synchronous handoff between the stdin reader and message processor.

See:

python-sdk/src/mcp/server/stdio.py

Lines 57 to 58 in 5439619

	read_stream_writer, read_stream = anyio.create_memory_object_stream(0)
	write_stream, write_stream_reader = anyio.create_memory_object_stream(0)

When the message processor is slow or blocked, the stdin reader cannot read new messages from stdin, making the server appear unresponsive.

Example Code

  import anyio
  import pytest


  @pytest.mark.anyio
  async def test_server_becomes_unresponsive_with_slow_processor():
      """Demonstrates how server becomes unresponsive during slow processing."""

      # Simulates stdio_server with default max_buffer_size=0 (synchronous handoff)
      send_stream, receive_stream = anyio.create_memory_object_stream(0)

      async def stdin_reader():
          # First message gets through
          await send_stream.send("request_1")

          # Second message (like a ping) blocks until first is fully processed
          await send_stream.send("ping")  # This will block for entire processing time!

      async def message_processor():
          # Process first message
          msg = await receive_stream.receive()

          # Simulate slow processing (database query, API call, etc.)
          await anyio.sleep(0.1)  # 100ms processing time

          # During this time, stdin_reader is completely blocked
          # No new messages (including pings) can be read!

          ping = await receive_stream.receive()  # Finally unblocks stdin_reader

      async with anyio.create_task_group() as tg:
          tg.start_soon(message_processor)
          await anyio.sleep(0.01)  # Let processor start waiting
          tg.start_soon(stdin_reader)

      send_stream.close()
      receive_stream.close()

Proposed Solution

Allow users to configure max_buffer_size > 0 to enable buffering, with a default value (0) that preserves the current behaviour.

async def stdio_server(
    stdin: anyio.AsyncFile[str] | None = None,
    stdout: anyio.AsyncFile[str] | None = None,
    max_buffer_size: int = 0, 
):
    # ...
    read_stream_writer, read_stream = anyio.create_memory_object_stream(max_buffer_size)
    write_stream, write_stream_reader = anyio.create_memory_object_stream(max_buffer_size)

Python & MCP Python SDK

• Python: 3.12
• MCP Python SDK: 1.12.4
• OS: macOS

Additional Context

The issue manifests in long-running servers where message processing can occasionally be slow (database queries, file operations, API calls).

The fix should make max_buffer_size configurable so users can add buffering to prevent the stdin reader from blocking during slow operations.

[yaowubarbara]

Hi, I'd like to take this on.

I saw PR #2163 was closed by the author — I'll pick up from where it left off.
My plan:

• Add configurable read_stream_buffer_size and write_stream_buffer_size params to stdio_server()
• Keep defaults at 0 for backward compatibility
• Add tests covering the buffered scenario (slow handler + concurrent requests)

Will submit a PR shortly.