MCPServer serves subscriptions/listen and advertises listChanged=true unconditionally — hold-open streams pin serverless invocations to the platform timeout

[tomoki-takahashi-oisix]

Summary

MCPServer always registers the subscriptions/listen handler (on_subscriptions_listen=ListenHandler(self._subscriptions)), and Server.get_capabilities() derives tools/prompts/resources.listChanged = true and resources.subscribe = true from the mere presence of that handler on the 2026-07-28 wire. A server that never publishes any change notification therefore advertises subscriptions, 2026-era clients open subscriptions/listen, and the response stream stays open until the client closes it (as specified).

On serverless HTTP platforms where a client disconnect does not propagate to the app (AWS Lambda Function URL with response streaming behind Lambda Web Adapter, in our case), every such stream pins one invocation until the platform timeout. There is no public way to opt out.

This is the same failure mode as #3492 (legacy GET stream), now on the modern wire.

Environment

• mcp 2.1.1 (local reproduction) and 2.2.0 (production, same behaviour)
• MCPServer(...) + mcp.streamable_http_app(streamable_http_path="/", stateless_http=True, transport_security=...), no json_response
• uvicorn 0.52.4 (uvloop + httptools), sse-starlette 3.4.8, Python 3.12
• AWS Lambda (container image, Function URL RESPONSE_STREAM, aws-lambda-web-adapter 0.9.1) — client disconnects never reach the ASGI app
• Clients: TypeScript SDK based (auto-open listen on connect) and a hosted connector

Reproduction (no Lambda needed)

import asyncio, json
from mcp.server.mcpserver import MCPServer

mcp = MCPServer("demo")

@mcp.tool()
async def hello() -> str:
    return "hi"

mcp.streamable_http_app(streamable_http_path="/", stateless_http=True)
sm = mcp.session_manager

META = {"io.modelcontextprotocol/protocolVersion": "2026-07-28",
        "io.modelcontextprotocol/clientInfo": {"name": "probe", "version": "1"},
        "io.modelcontextprotocol/clientCapabilities": {}}

def scope(body, method):
    return {"type": "http", "http_version": "1.1", "method": "POST", "scheme": "http", "path": "/",
            "raw_path": b"/", "query_string": b"", "server": ("localhost", 3000), "client": ("127.0.0.1", 1),
            "headers": [(b"host", b"localhost"), (b"accept", b"application/json, text/event-stream"),
                        (b"content-type", b"application/json"), (b"content-length", str(len(body)).encode()),
                        (b"mcp-protocol-version", b"2026-07-28"), (b"mcp-method", method.encode())]}

async def post(msg, timeout=3):
    body = json.dumps(msg).encode(); got = False
    async def receive():
        nonlocal got
        if not got:
            got = True
            return {"type": "http.request", "body": body, "more_body": False}
        await asyncio.sleep(3600)          # like Lambda: no http.disconnect ever arrives
    events, done = [], asyncio.Event()
    async def send(m):
        events.append(m)
        if m["type"] == "http.response.body" and not m.get("more_body", False):
            done.set()
    task = asyncio.create_task(sm.handle_request(scope(body, msg["method"]), receive, send))
    try:
        await asyncio.wait_for(done.wait(), timeout); print(msg["method"], "completed")
    except asyncio.TimeoutError:
        print(msg["method"], "STILL OPEN after", timeout, "s")
    print("  ", b"".join(e.get("body", b"") for e in events if e["type"] == "http.response.body")[:200])
    task.cancel()

async def main():
    async with sm.run():
        await post({"jsonrpc": "2.0", "id": "d1", "method": "server/discover", "params": {"_meta": META}})
        await post({"jsonrpc": "2.0", "id": "listen:1", "method": "subscriptions/listen",
                    "params": {"_meta": META, "notifications": {"toolsListChanged": True}}})

asyncio.run(main())

Output:

server/discover completed
   {"jsonrpc":"2.0","id":"d1","result":{"cacheScope":"private","capabilities":{"prompts":{"listChanged":true},"resources":{"listChanged":true,"subscribe":true},"tools":{"listChanged":true}}, ...
subscriptions/listen STILL OPEN after 3 s
   event: message\r\ndata: {"jsonrpc":"2.0","method":"notifications/subscriptions/acknowledged","params":{"_meta":{"io.modelcontextprotocol/subscriptionId":"listen:1"},"notifications":{"toolsListChanged":true}}}

The server has nothing it could ever publish, yet it advertises listChanged: true and holds the listen stream open.

Impact observed in production

After migrating from mcp 1.29 (FastMCP) to 2.x, the only method that never completed was subscriptions/listen: in a 30-minute sample, 19 of 19 listen POSTs were still unanswered after 60 s while every other method (tools/call, tools/list, prompts/list, resources/list, server/discover, initialize) completed in milliseconds. Each open stream ran to the Lambda timeout (900 s); ~1,300–1,800 such invocations per day accounted for 99% of the function's billed GB-seconds. Clients re-sent the listen request immediately after each timeout, so the streams were permanently occupied.

Current workaround

mcp._lowlevel_server._request_handlers.pop("subscriptions/listen")

With the handler gone, get_capabilities() reports listChanged: false / subscribe: false, and a listen POST gets the SDK's standard 404 + JSON-RPC -32601, which the TypeScript client treats as a soft error ("a failed auto-open MUST NOT fail connect"). It works, but it depends on two private attributes.

Request

A public way to run a server without subscriptions/listen, for example MCPServer(subscriptions=None) meaning "not served" (today None means "use the in-memory bus"), or an explicit listen=False / serve_subscriptions=False flag — and have get_capabilities() derive listChanged=false from it, as it already does from handler presence. Alternatively, a documented guidance for stateless/serverless deployments.

Thanks for the SDK — the 2.x transport is otherwise working well for us.

[owlofjaylan]

yea annoying ash fr i reported this too #3357
@maintainers when will this be fixed? happy to implement a fix myself and draft a pr but guidelines say the issue has to be assigned to pr author by a maintainer or carry the help wanted label.

[mikerva]

I had this exact issue yesterday as well. Bump! ^^
@owlofjaylan