0xreversed/with.ai
AppsPartialSource available
Target: Virtual Desktop Streamer (Windows and macOS)

Ported the Virtual Desktop streamer to Linux by having Claude Code analyse the installers

Virtual Desktop Streamer (Windows and macOS): Quest streaming from Linux, rebuilt from the vendor own protocol libraries.
Image: Reversed with AI

What was done

cudanexus port reverse engineered both official Virtual Desktop installers with REA and Claude Opus, reuses the macOS streamer protocol libraries and replaces the Mac layer with Linux capture, encoding and input.

Virtual Desktop has no Linux streamer, so headset users of Linux were stuck. cudanexus port analysed the two official installers with Claude Opus working through Claude Code and the REA toolkit, from a single prompt plus one fix prompt, and rebuilt the macOS streamer logic for Linux. The repository contains no vendor code: the build extracts the protocol libraries from the official installer you download, patches only the three macOS specific calls and keeps built packages for your own use.

How it was done

The Windows streamer, a .NET Framework build with native drivers, cannot be ported, so the macOS one, a .NET 8 app whose protocol libraries are platform neutral, is the base. The port reuses those vendor libraries for discovery, pairing, remote connections and encryption, and replaces the Mac layer with Linux capture through xdg-desktop-portal or X11, GStreamer encoders from NVENC and VA-API to x264, PulseAudio and PipeWire audio, uinput keyboard, mouse and pointer devices, and a local web page and command line interface for settings. Mock headset tests cover seven scenarios end to end against the installed packages.

Results

Version 1.1 adds socket pacing against Wi-Fi queueing that measured 258 to 337 ms delays down to 1 to 24 ms in tests, network change detection that makes the PC findable again in seconds, adaptive bitrate and a scroll speed recovered from the Windows streamer input service by decompilation. The packages install from Ubuntu 20.04 through Fedora and openSUSE. The original tester ran no real headset, so a work in progress tag stays on until real use confirms the feature set.

From the source