Streaming HTTP - how do you *actually* stream tool responses?

[bendavis78]

With the release of the Streaming HTTP protocol, I fully expected to be able to do this:

@mcp.tool()
def get_content(input: str) -> AsyncGenerator[str]:
    for foo in bar():
        yield foo.content

So far I'm only seeing examples where I can return a value. If I try the above, the client just receives:

""<async_generator object get_content at 0x78ba7a28d930>""

[fdurant]

Exactly the same question here! I have tried to fiddle with the Context, but it crucially lacks a send_content method. Tragic, because it already sends intermediate progress via report_progress.

[jlowin]

"Streaming" HTTP is misleadingly named... it doesn't actually stream responses (at least at this time), it's just a persistent connection over which many request/response pairs can be sent. I hope that streamed responses are added to the spec at some point, but at this time tools return a single payload (at least according to protocol -- it seems like such a clear win that I'd love to find a way to support this as a superset capability)

[bendavis78]

@jlowin Isn't SSE capable of doing that? Why would they be replacing SSE with Streamable HTTP if that will no longer be possible?

[jlowin]

Both SSE and Streaming HTTP are capable of streaming responses from a server. It's not a question of technological feasibility, it's simply unsupported by the MCP protocol spec, which is why I'd be interested in supporting it in FastMCP (but would only work with fastmcp servers / clients unless we overload another part of the spec like logging, but could create unexpected behavior in non-FastMCP clients)

[bendavis78]

Hmm, I read into the spec and realized they aren't actually replacing/deprecating SSE (that was confusing), it seems like they're just moving it into the functionality of the Streamable HTTP transport...

In the Streamable HTTP transport, the server operates as an independent process that can handle multiple client connections. This transport uses HTTP POST and GET requests. Server can optionally make use of Server-Sent Events (SSE) to stream multiple server messages. This permits basic MCP servers, as well as more feature-rich servers supporting streaming and server-to-client notifications and requests.

I may be confused, but that makes it sound like streaming responses are still part of the spec.