* feat(xbox): install legacy EAppx titles through Windows Update A legacy Store title does not ship as MSIXVC. It ships as an encrypted EAppx bundle, which streaming rejects, so the store could list the game and never install it. Accept EAPPX and EAPPXBUNDLE beside MSIXVC, carry the format through the catalogue, the install record and the cached library, and dispatch legacy-install for those formats. Two backend patches do the work: - droiddeck-legacy-packages.patch lets the plan, license and extract verbs handle a legacy package, and expands a bundle in place (nesting bounded at three), decrypting each inner EAppx with the same keys. - droiddeck-legacy-install.patch adds legacy-install: resolve, license, download, extract, hoist AppxManifest.xml to the game root. Resolution asks the Xbox content service first and Windows Update second, because some legacy products return no files from the content service at all. The WU endpoint needs no Xbox credential but chains to Microsoft's ECC Product Root CA 2018, which Android's CA bundle does not carry, so that one public root is pinned for those requests. The manifest reader accepts AppxManifest.xml as well as MicrosoftGame.config, since a legacy package carries only the former. Running it against a real title exposed four things that only show up at full size: - A licence names a key by its 16-byte content id, while a bundle header names it as a (content id, package id) pair. Reading the 16-byte id yields a different pair, so a complete key file reported every key missing and no bundle could be opened. Keys are now aligned against the header's own ids before the header is judged. - Windows Update answers with several updates whose PackageContentId differs from the bundle's own. Licensing used to take the first and fail with content-http 400; every candidate is now tried in order, preferring the one the container is named after. - A container already in .xodus-container is reused rather than fetched again, so a failed run does not cost another download of tens of gigabytes. The space preflight no longer counts a container twice for the same reason. - A bundle that carries add-on packages unpacks each into its own folder, leaving the manifest one level down. Hoisting lifts the largest folder that holds a manifest, so the game still lands at the root beside its add-ons. Unpacking now reports bytes written, so the install bubble fills while a package is decrypted instead of holding a single unchanging word. * fix(xbox): unpack a legacy package across cores and report the bytes Unpacking one title ran as a single thread on a saturated core at about 40 MB/s. Files sit at independent offsets, so each worker now opens its own handle and takes every Nth file, with a 1 MiB read buffer in place of the 8 KiB default that turned one 64 KiB block into eight reads. Measured on the Forza Horizon 4 bundle, writes went from 40 MB/s to about 200 MB/s. Two things the first full-size run exposed: - The container was deleted only after the whole nested tree was unpacked, so a title that holds its package as one member held the archive and the decrypted copy at once. On this device that runs out of space partway through the second stage. The container is now deleted once its members are on disk, which frees the archive before they are decrypted. - A bundle member is written as a single file, so a counter that advances when a file finishes stays at zero for the whole of the longest stage. The reported figure now also follows the growth of the destination while that is the case, and never moves backwards when the container is freed. Progress is exact during the per-file stage: each file's declared length is added as it lands, instead of walking the whole destination every tick. * docs(xbox): record the completed legacy install The evidence section said a finished legacy install and its decrypted contents were unverified, because the first run was still unpacking when it was written. That install has now completed on the same device: the 85,219,173,025-byte container produced 79.3 GB at the game root with AppxManifest.xml and the package DLLs directly beneath it, and the launcher lists the title as installed. It also records what unpacking costs, since that decided the change: one thread held a core at 100% for about 40 MB/s and four reach about 200 MB/s, and releasing each container once its members are on disk is what keeps the peak footprint at one package rather than the archive plus its contents. The launch is not part of this evidence, and the entry says so. * fix(xbox): place the 64-bit framework beside a legacy title A Windows Update framework category lists one package per architecture under identifier-only names. The first package was unpacked, which on this device was the x86 build, so a 64-bit title got 32-bit DLLs and every load failed with STATUS_INVALID_IMAGE_FORMAT. Candidates are unpacked now and only the one whose manifest declares ProcessorArchitecture="x64" is kept. The framework folder was also never created, so a finished download failed at the move with "no such file or directory". unzip.rs belongs to this patch: legacy-install calls it and the recipe left it out, so a rebuild from the pinned source could not compile. * feat(xbox): report a UWP entry point at launch An unpacked package can name a COM class as its entry point instead of Windows.FullTrustApplication. That class is a UWP application, which expects the app model to activate a view; Wine has no host for it, so the launch died inside the loader with no explanation. The launch now reads the entry point and reports uwp-entry-point, and full-trust packages are unaffected. Forza Horizon 4 is such a package: its manifest names Forza.App. Its frameworks are installed and the loader is clean before that point. * fix(xbox): report why a launch was refused A failed prepare replaced the backend message with a fixed sign-in hint, so a title that was not installed or a runtime that refused all read as a bad account. The reason is kept, with paths and URLs dropped, and written to the store log. * fix(adb): dismiss the keyboard on Search and Go Search and Go were terminal IME actions but were not handled, so focus stayed in the field and the keyboard covered the results it had just fetched. * fix(xbox): keep a cache row without a legacy format plain Decoding added a format key with an empty value to every cached title, so a library restored from storage no longer matched the one that was saved. * test(ui): match the manifest message a legacy package reports The message covers both a MicrosoftGame.config and an AppxManifest, so the assertion follows the text the content dialog now prints. * fix(xbox): refuse to start a title that needs the Windows app model A legacy package that names an entry point other than Windows.FullTrustApplication is a UWP application. Wine has no app-model host, so starting its executable crashes inside the loader and the launch looks like nothing happened. Read the entry point from the manifest the installer already parses and stop the Play action with a message instead. Forza Horizon 4 names Forza.App, so it installs and reports as installed but cannot be started here yet. * fix(xbox): dispatch on the container format of the chosen package The catalogue records the format of the first installable desktop package, and the picker only replaced the content id. An offer that lists MSIXVC and a legacy container would then hand the legacy id to streaming, or the streamed id to legacy-install, and the install failed after resolution had succeeded. The choice now carries its own format and the catalogue value only fills a gap. * fix(xbox): keep an extraction reason that the sanitizer rewrote The extraction sanitizer replaces anything outside letters, digits, space, dot, underscore and dash with a question mark, so a reason such as "(os error 28)" arrived as "?os error 28?" and the allowlist dropped the line. The failure then showed only the stage code, with no cause. Add that replacement character to the class the allowlist accepts. * fix(xbox): print the reason a launch was refused The account channel sends a reason with a refusal, and the guest replaced it with a fixed sign-in message, so a title that was never installed still read as a bad account. The guest now shows the reason it was given, and a token holding a path is dropped whole instead of having its slashes removed and its directory names kept. * build(xbox): leave RUST_TEST_THREADS unset when the host has no value The recipe always passed the variable, empty when the host had none, and the pinned test harness rejects an empty count. On an ARM64 host that runs cargo test, so the default local build failed before compiling. * fix(xbox): rebuild the backend patches from the tree that ran The two patches in this branch were not the ones the verified install ran, and one of them did not compile. `unzip.rs` inflates deflated entries but the dependency that provides the inflater was never added, so `prepare.py` produced a tree that failed at `flate2::read::DeflateDecoder`. Anything else missing fell into the same gap: the install stage reported no bytes, so the bar held one value for the whole unpack, and a container a stopped run had finished was downloaded again. Rebuilt from the patched source, and fixed the defects a review of the patch found: - Unpacking reports bytes again: the workers count each file's declared length, and for the single-file bundle member the folder's growth stands in, which is what the device run showed. - A finished container is reused, so a retry after a failure past the download does not spend hours fetching it again. The space preflight counts it as reclaimable, since the retry truncates it first. - An install over a folder that already has a manifest replaces the old files instead of leaving the new package nested and reporting success. A first install still refuses to merge into an existing target. - The manifest search looks at the depth a nested bundle writes, not only one level down. - `withdraw_libraries` removes only the DLLs `copy_libraries` recorded copying, so a framework replacement cannot delete a library the title owns. - ZIP64 sizes and offsets stay `u64`; truncating them to `u32` read the wrong local header past 4 GiB. - The extraction checks the blockmap length against the footer's before writing, which the crate only does when asked. - The container being unpacked is never treated as a nested member, so `extract-eappx title.eappx .` no longer re-enters and deletes its own input. - `flate2` is declared and the lock is updated. `prepare.py` reproduces this from a pristine checkout and `cargo test` passes on it. * fix(xbox): apply the legacy install patch over the Nexus 0.2.15 backend * fix(xbox): install only the newest x64 build of a legacy title and size the space check at its real peak * fix(xbox): install a store package that is a plain archive instead of an XVD * fix(xbox): refuse a package entry that ends before its recorded size * fix(xbox): credit only complete staged containers toward the free space a legacy install needs * fix(xbox): bound each package entry to its data and check ZIP64 decoded sizes * fix(xbox): refuse duplicate package names and clear an unfinished download before measuring space * fix(xbox): bound package entries by the next header and stage a plain package until it validates * fix(xbox): hold the install folder lock through a legacy install * fix(xbox): fall back to an older build when Microsoft refuses the newest one a license * fix(xbox): keep the registered device across sign-out * feat(xbox): explain device-limit and license refusals, with a link to manage devices * fix(xbox): locate each build only once licensed, roll back a failed package swap, report the final license answer * fix(xbox): skip trial and preorder packages, read runtime behavior, keep licensing timeouts * fix(xbox): keep a failed package swap recoverable, size bundles up front, report key-write and framework errors cleanly * fix(xbox): drop a pending manifest check on any newer launch, keep committed backups out of restores, match the app-model order * fix(xbox): cancel pending launch checks on a new intent, remember entries a rollback could not remove * fix(xbox): clear the removal list before a restore moves old files back --------- Co-authored-by: TimmyCraft <46533729+TimmyTurner51@users.noreply.github.com>
DroidDeck brings the SteamOS experience to Android: Steam Gaming Mode, full Linux desktop, integration with GOG, Epic Games, Xbox, and other storefronts, Steam VR, and more!
100% open source. No ads. No telemetry. No proprietary components. No source withheld.
The official website is droid-deck.dev. Only download DroidDeck from this repository's Releases.
Install
- Android 9 or newer on Adreno 730 or newer, or any 8xx. Adreno 6xx, Mali, Xclipse and PowerVR are experimental. Adreno 710 is currently unsupported. No root needed.
- About 3 GB free for the runtime, plus 1.1 GB for the desktop and emulators.
- Install the APK, install the Linux runtime, press Play and sign in to Steam.
- On Android 14 and newer, turn on Disable child process restrictions in Developer options. On Android 12 and 13, Steam's first launch shows a Fix it for me button that walks you through it.
Smart glasses setup is in the support matrix. Release notes are in docs/releases.
Help, Preview builds and device reports are on Discord. For bugs, post in #bug-reports with the zip from Share logs.
Build
Needs Docker, Java 17, the Android SDK and NDK, Rust via rustup, Go, a host C compiler, the x86_64 and i686 MinGW-w64 compilers (i686 with DWARF-2 exceptions, as Debian and Ubuntu ship it; SJLJ builds such as Homebrew's cannot link Rust's i686-pc-windows-gnu), GitHub CLI, Python 3 and zstd. On x86_64 Linux, Docker also needs an arm64 binfmt handler (docker run --privileged --rm tonistiigi/binfmt --install arm64).
tools/build_local.sh
The APK lands in app/build/outputs/apk/release/. Install it with tools/deploy_local.sh. After the first run, ./gradlew :app:assembleRelease reuses the prepared inputs. Unit tests run with ./gradlew -PskipRust=true :app:testDebugUnitTest, and the helper and preload suite with bash tools/test_local.sh.
On Windows, see tools/run_windows.ps1 and tools/prepare_windows_runtime.py.
Limits
Compatibility and performance vary by device. Desktop compositing is software rendered, and Firefox sandboxing is reduced under proot.
Credits and licence
DroidDeck is GPL-3.0. Steam and Proton belong to Valve Corporation. VITURE, RayNeo, XREAL, Rokid, REDMAGIC, Nintendo Switch and Joy-Con are trademarks of their owners. This project is not affiliated with any of them.
Runtime, Wine and Proton
- Runtime, shim, input and controller work build on WinNative and Bannerlator by maxjivi05.
- mmap ring buffer controller input by ribbit384 (WinNative #345).
- FEX (MIT) runs x86 code. DroidDeck's build puts the platform name under
AT_PLATFORMso sharun AppImages by VHSgunzo (MIT) start through its userland-execve loader (MIT), and its GL thunk hands X errors to the program's own error handler. FEX's code disk cache was checked in its source. - x86 AppImages, and arm64 ones whose runtime can't unpack on Android, are unpacked with uruntime by VHSgunzo (MIT), shipped unmodified with its licence.
- x86 programs linking OpenSSL 3 use the copy in the Steam Linux Runtime 4.0. x86 Linux games get the
/etc/asound.confthat pressure-vessel writes. WriteProcessMemoryhooks fall back to/proc/<pid>/memthe way Wine'sserver/ptrace.cdid before Wine 11.- Windows programs on noexec storage get executable copies of their image pages, following @yujikaido's finding in droiddeck-rp4pro that Wine's own processes need the noexec shim.
- Wine write watches use userfaultfd support in Proton's
dlls/ntdll/unix/virtual.c(LGPL-2.1-or-later). libddsession supplies the arm64 syscall number and answersPAGEMAP_SCANfrom/proc/self/pagemap. - libddsession runs Wine's arm64
SIGUSR2context handler (dlls/ntdll/unix/signal_arm64.c, LGPL-2.1-or-later) in place on a frame laid out like Linux'sarch/arm64/kernel/signal.c, and servesfill_vm_countersindlls/ntdll/unix/process.cthe kernel's own/proc/<pid>/status. - droiddeck-ntsync follows the Linux ntsync semantics in
Documentation/userspace-api/ntsync.rst(GPL-2.0). - The proot fast path orders bindings like PRoot's
path/binding.c(GPL-2.0-or-later). - DiRT 3 and DiRT Showdown keep their telemetry socket the way RoccoC's DiRTTelemetryErrorFix (MIT) does, based on the fix by Carsten Gottschling.
- Zink on PowerVR keeps colour textures in colour with Icenowy Zheng's upstream Mesa fix d55418dae7 (MIT), carried as a runtime patch.
- Games and AppImage, Flatpak and Snap programs such as emulators use
TU_DEBUG=sysmemby default, as does the Steam client. Turnip sysmem rendering can be turned off in Performance settings, except with the Adreno 710/720 driver, which always uses it. - Black Ops III and Dark Souls Remastered start without umu-protonfixes (BSD-2-Clause), whose
311210.pyand570940.pyfail under ARM64 Proton. - ReSkate (GPL-3.0) by Dingo Shenanigans uses the launch option from its README and D3D12 feature level 12_1.
- The arm64 client's sign in was checked by reading Steam's saved token with the
ConnectCacheformat from deadlock-api-ingest.
Stores
- Epic Games and GOG build on Phobos's adaptation of Legendary and gogdl. GOG install script registry entries are read the way Heroic Games Launcher (GPL-3.0) reads them. GOG cloud storage requests send the Galaxy communication-service User-Agent that gogdl and Lutris (GPL-3.0) send.
- Epic's official Linux launcher preview runs through FEX. That setup is outside Epic's listed Linux support.
- The launcher's .deb gets the WebKitGTK and GTK 3 libraries it needs from unmodified Debian 13 packages, following @zooluis's research in #2268.
- Battle.net is by @kayakyakr. It runs Blizzard's own client in Proton (downloaded from Blizzard, not included) and shows key art from the Battle.net Shop. Without Steam, its games see the pad as an Xbox 360 controller via
PROTON_SPOOF_STEAMINPUT_VIDPIDin Proton'sdlls/winebus.sys. - The EA store follows the EA app's protocols as documented by ArmchairDevelopers' Maxima (GPL-3.0): sign-in, the GraphQL queries (included verbatim, since EA matches them by hash), build archives, online activation licences and the LSX game endpoint. The code is DroidDeck's own.
- Battle.net installs drive Blizzard's Agent the way @barncastle's Battle.Net-Installer does, with requests from wowdev.wiki. Read without taking code: @mubaraknumann's unifideck, @djmittens's neo-mittens
bnetswitch, @BCD1210's soju and Lutris'sproduct.dbreader. - The Rockstar launcher service looks for launcher pipe names in the executables beside the game, since CitizenFX's
ros-patches-fiveshows GTA V openingGTAVLauncher_Pipeand @pingod's rdr2-linux-launcher notes that Rockstar's protected executables resolve such strings only at run time. Read without taking code. - World of Warcraft addon imports replace an addon's folder from a temporary copy, as WowUp does, and read flavour folders the way WoWDeck does. No code from either is used.
- The World of Warcraft addon Download window searches WoWInterface through its MMOUI API. CurseForge opens in the user's browser, with no API and no key.
- The Xbox store runs @domsvoboda's nexus-backend 0.2.15 (GPL-3.0, a fork of Xodus), and Xbox games load the Windows modules of his nexus-runtime (LGPL-2.1-or-later), including the C++ nexus-xgameruntime and
wintypes.dll. Content licensing sends delegation-compact device tickets in the licensing wire form, as nexus-backend's Store license path (commit27c335e) does. Minecraft's GameInput setup was checked against @Wyze3306's BedrockOnLinux (MIT). - The Xbox Cloud Gaming player uses the WebRTC client of @unknownskl's Greenlight (MIT), and its offer tweaks and sharpening follow @redphx's Better xCloud (MIT). See
tools/xcloud-player/NOTICE.md. - The itch.io butler pack ships butler's
7z.so, itch.io's build of 7-Zip by Igor Pavlov (LGPL-2.1, unRAR restriction on RAR code), unmodified. Other archives use Apache Commons Compress and the runtime's bsdtar (libarchive). - The Store uses Flatpak (LGPL-2.1) through DroidDeck's own bubblewrap shim. Chromium and Electron apps start directly, since zypak by @refi64 (BSD-3-Clause) always uses Flatpak's spawn portal.
- Folder sync runs Syncthing (MPL-2.0), built from @researchxxl's syncthing-android fork (MPL-2.0).
Audio, video and HUD
- Steam audio goes through @The412Banner's DirectAudio PulseAudio sink (LGPL-2.1-or-later), built from
tools/directaudio/sink. - The DroidDeck HUD is @The412Banner's FusionHUD (GPL-3.0 with an attribution term). Its readings match Valve's MangoHud build (MIT), with frame times from gamescope patch 0125.
- The HUD's CPU, GPU load, thermal and battery readers come from @xXJSONDeruloXx's work.
- LSFG frame generation comes from Camille LaVey and the Eden project, following lsfg-vk, ported by @maxjivi05. It needs your own copy of Lossless Scaling and ships none of its shaders.
- Black frame insertion follows the design of @artzox's BFI Presenter (GPL-3.0): the strobe runs on the display's refresh clock and the game is paced to the lit refresh. Both build on Blur Busters' research on MPRT and motion blur. No BFI Presenter code is used.
- Remote Play decodes through Android's MediaCodec or the FFmpeg (LGPL-2.1-or-later) in Steam's runtime. Qualcomm low latency settings follow Moonlight (GPL-3.0).
- PipeWire (MIT) sockets use the settings in its
module-access.c, and WirePlumber (MIT) keeps those clients at full permissions.
Desktop
- The KDE pacman repository snapshots Arch Linux ARM and installs with pacman (GPL-2.0-or-later).
- KWin (GPL-2.0-or-later) is patched to pass on the scroll data its
wayland_backend.cppleaves as a TODO. - Read without taking code: Qt's
qwaylandinputdevice.cppandqguiapplication.cpp, Konsole'sTerminalDisplay.cpp, and Mesa'su_threaded_context.candzink_resource.c. - The Steam Keyboard window rule follows Valve's
steamdeck-kde-presetswith Bazzite's0001-steam-virtual-keyboard.patch. - Steam's Browse buttons open a GTK 3 file chooser via PyGObject (LGPL-2.1-or-later), shown over Steam by gamescope (BSD-2-Clause) through
GetAppIdFromReaper. - The AYN Odin 2 phantom second display was confirmed by @SapphireRhodonite's WatermelonDS 0.0.5 notes and iiSU #363, without taking their code.
Smart glasses
- Builds on WinNative's VITURE work. Protocol checked against @wheaney's XRLinuxDriver, @badicsalex's ar-drivers-rs and go-macos/iokit. The optional VITURE SDK is not included.
- The VITURE protocol V2 frames adapt @elasticjava's viture-v2 (MIT) by Holger Bartnick. See
docs/third-party/viture-v2. - RayNeo adapts @informalTechCode's RayNeo-Air-Series-3DoF-Scaffold (Apache-2.0) and references @verncat's RayNeo-Air-3S-Pro-OpenVR (MIT) and Monado (BSL-1.0). See docs/third-party/rayneo.
- XREAL 1S references XRLinuxDriver and xrealOneDeviceKit by @wheaney, @rohitsangwan01's xreal_one_driver (MIT) and @0xcaff's xr-tools.
- Rokid follows ar-drivers-rs, @badicsalex's Void Computing write-ups and Monado's Rokid driver.
- REDMAGIC display findings draw on @mekhontsev's MagicDesk investigation and @Coding-BR's decompiled Nubia kernel.
VR, Joy-Con and hand tracking
- Joy-Con VR ports the arm model from Google's GVR SDK for Unity (Apache-2.0) and adapts the hand skeleton from Valve's OpenVR
handskeletonsimulationsample (BSD-3-Clause), which follows Moshi Turner's work in Monado. Index offsets come from Monado'svive_poses.c(BSL-1.0). - Joy-Con research: Linux's hid-nintendo by Daniel J. Ogorchock, @dekuNukem's Nintendo_Switch_Reverse_Engineering, SDL and AOSP. Checked against ALVR, OpenComposite, PSMoveSteamVRBridge, @F8LEFT's JoyConVr, @krnelpan1c's Daydream-Controller-SteamVR, @yoyobuae's joyconvr, Moonlight and Dolphin. See docs/third-party/joycon.
- Camera hand tracking runs Google's MediaPipe Pose Landmarker (Lite), BlazePose GHUM 3D and Hand Landmarker (Apache-2.0) on CameraX frames (Apache-2.0), with usage reports off. Checked against JoyConVr's camera app (MIT).
- Tracking math: the GVR arm model's accelerometer mode, Eric Foxlin's zero velocity updates, Frosio, Pedersini and Borghese's accelerometer autocalibration, a Kalman filter (Kalman, 1960), Horn's quaternion method (1987) and the 1€ filter (Casiez, Roussel and Vogel, 2012).
- Gestures follow ALVR's
hand_gestures.rs(MIT) and Meta Quest's menu pinch, tuned on HaGRID photos (CC BY-SA 4.0, not included). - The glove model follows OpenVR's render model reference, is meshed with surface nets (Gibson, 1998; Mikola Lysenko), and uses Inigo Quilez's smooth minimum, distance field ambient occlusion and polygon distance (MIT). See docs/third-party/handtracking.




