It's a library that is linked to in place of an operating system - so whatever interface the OS provided (syscalls+ioctls, SMC methods, etc.) ends up linked / compiled into the application directly, and the "external interface" of the application becomes something different.

This is how most unikernels work; the "OS" is linked directly into the application's address space and the "external interface" becomes either hardware access or hypercalls.

Wine is also arguably a form of "library OS," for example (although it goes deeper than the most strict definition by also re-implementing a lot of the userland libraries).

So for example with this project, you could take a Linux application's codebase, recompile it linked to LiteBox, and run it on SEV-SNP. Or take an OP-TEE TA, link it to LiteBox, and run it on Linux.

The notable thing here is that it tries to cut the interface in the middle down to an intermediate representation that's supposed to be sandbox-able - ie, instead of auditing and limiting hundreds of POSIX syscalls like you might with a traditional kernel capabilities system, you're supposed to be able to control access to just a few primitives that they're condensed down to in the middle.

> So for example with this project, you could take a Linux application's codebase, recompile it linked to LiteBox

If you have to recompile, you might as well choose to recompile to WASM+WASI. The sandboxing story here is excellent due to its web origins. I thought the point of LiteBox is that recompilation isn’t needed.