Support Stateless MCP Server Connections

[BrandonShar]

Description

With the streaming http protocol, MCP servers now have the option to be stateless and not care about long-running sessions. If we don't need long-running session support, that also allows us to skip the initial connection to the MCP server that occurs in agent.run_mcp_servers() which is a little performance boost and reduces start-up coupling with any MCP servers.

I'm handling this in a project right now by overriding the MCPServerHTTP's aenter method and running everything except for the client initialize and set logging methods.

I'm happy to put up a PR for this, but I wanted to see if you'd have interest in the idea first. I think the simplest way to handle it would be an additional argument on the MCPServerHTTP class called something like stateless or maybe something more obvious like skip_initial_connection.

References

No response

[DouweM]

@BrandonShar Do I understand correctly that we'd basically always want this with Streamable HTTP servers, and we just need the current eager behavior for SSE servers? Ideally we'd be able to detect which the user is doing.

[BrandonShar]

I think we'd still need to support the initalization for protocol reasons (but maybe we could put off the initialization until the first request?). Streamable HTTP servers can optionally support the concept of "state" via a persistent session id that's passed each time by the client.

The protocol says:

A server using the Streamable HTTP transport MAY assign a session ID at initialization time, by including it in an Mcp-Session-Id header on the HTTP response containing the InitializeResult.

That reads to me that they still expect an SSE style initialization request when session ID's are involved, but it's unclear to me if that's technically necessary (or if we could hold off and send it directly before the first request).

I do wonder if, in practice, anyone is actually going to use that. It feels like stateless HTTP is a more reasonable default and long-running sessions would be better managed by some type of auth.

[maxonics]

I tried to use a mcp server in stateless mode and indeed it doesn't work. Would be great to support it 🙂

[DouweM]

@BrandonShar I haven't read up on the details here, but nothing you write sounds unreasonable and the demand is clearly there (hi @maxoniboo!) so if you could submit a PR and call out any decisions you made that we could go another direction on, that'd be great! I find it easier to reason about real code than protocol specs :)

[BrandonShar]

I find it easier to reason about real code than protocol specs

You and me both! 😄

Sounds good, I'll get something up in the next day or two.

also, @maxoniboo did you notice a specific error? It should work fine with stateless servers, it currently just makes some unnecessary extra calls for them.