Nothing about the RSC format is privileged in the React client (i.e., react + react-dom). You can always make your own serialization format as long as it can represent React elements and promises somehow. You can of course also use the RSC wire format and pin to a specific client version (or vendor/fork it) but it’s not a stable format today because it’s not clear if many people would use one and we want to leave room to improve the format as new optimization opportunities arise.
Though you do lose a lot of the magic if you don’t have JS on the server at all. Part of the core proposition of RSC is that it’s easy to move logic across the boundary as your app evolves. What goes over the wire can be a performance decision instead of a “that’s too much of a pain to move to the other side I’ll just leave it” decision. For that reason I’d generally make a thin “backend for frontend” in JS that talks to an Erlang/Rust/etc backend. It’s conceptually part of the frontend but runs on your server.
That's fair, though IMHO some marketing from Vercel does make it difficult for someone not versed in the internals to understand what is "privileged in the React client" and what is not. That's their prerogative as a leading maintainer, of course.
And completely understand that freezing a format causes all kinds of difficulties, and semver alone doesn't solve for the resulting community frustration. So I understand why the decisions were made - it's just all a bit foreign to me compared to the groundbreaking simplicity of React's APIs in its earliest days, where if you understood a render cycle the entire world was your oyster. The internals of the renderer were always fine to be "off limits" - but wire formats just feel quite different to me.
(Also, just want to say you're an absolute hero of mine, and my hopefully-gentle nudging on this is because I love everything the React team has created, and I want React to see its full potential!)
It’s fair, the docs could be better! And there are certainly subtleties with hooking up the router and bundler and everything. However I will say that really there is no funny business going on.
Basically the entire elaborate RSC apparatus just ultimately outputs a tree of plain React elements with promises and <Suspense> to make it streamy. Sounds like groundbreaking simplicity to me! I kid a bit but really I’m not sure that the RSC wire format is any easier as an output target than the existing public React client API.
Maybe I should make a little demo to show that.
But anyway the point is that the wire format is an implementation detail because the two sides of the wire are actually the same program. Even if they’re running on two different computers.
Though you do lose a lot of the magic if you don’t have JS on the server at all. Part of the core proposition of RSC is that it’s easy to move logic across the boundary as your app evolves. What goes over the wire can be a performance decision instead of a “that’s too much of a pain to move to the other side I’ll just leave it” decision. For that reason I’d generally make a thin “backend for frontend” in JS that talks to an Erlang/Rust/etc backend. It’s conceptually part of the frontend but runs on your server.