Deadlock Bug

This is reproducible in streamable-http and sse but not stdio. That makes me think it's a core concurrency issue in mcp.

Running

In all these examples, change the numeric argument on the client. What I've found

• 10 — works easily, fast
• 100 — works
• 200 — works in streamable-http, deadlock in sse
• 250 - deadlock in streamable-http
• 500 — deadlock
• 1000 — big deadlock

For both streamable-http and sse, the following error occurs frequently, followed by the script hanging:

Error in post_writer: All connection attempts failed

If it fails and I kill the client but not the server and re-run with a much lower number (e.g. 250 and then 10), the second small run works fine.

conclusion: maybe the deadlock is in the client and the server is uneffected?

The --tool-concurrency flag changes the concurrency to not include connecting, one single connection serves all threads and only a tool call is made. In my experiments, this doesn't trigger a deadlock issue.

Running streamable-http

server:

uv run --script server.py streamable-http

client:

uv run --script client.py 0.0.0.0 8000 streamable-http 1000

Running sse

server:

uv run --script server.py streamable-http

client:

uv run --script client.py 0.0.0.0 8000 sse 1000

Running stdio

# lots of excess noise on stderr
uv run --script client.py 0.0.0.0 8000 stdio 10 2>/dev/null