Install Apollia Desktop on Windows

The application is Apollia Desktop. Its installers and the files they put on disk still carry the name of the engine underneath, Apollia OS, which is the name you will read below.

Every file below is attached to each release on the publication page, https://github.com/Apollia-OS/apollia-os/releases.

• Apollia-OS_<version>_x64_en-US.msi (recommended): standard Windows installer with Start menu entries + uninstaller.
• Apollia-OS_<version>_x64-setup.exe: the NSIS installer. A single file, and it installs rather than runs in place: it registers an uninstaller in Settings > Apps exactly as the .msi does.
• apollia-os-windows-x86-cpu.zip and apollia-os-windows-x86-vulkan.zip: the two command-line bundles, one per engine, CPU or Vulkan.

Requirements

• Windows 10 22H2 / Windows 11.
• 4 GB of free RAM minimum.
• For GPU inference: a Vulkan-capable driver. The accelerated engine for NVIDIA cards, which the application offers after install, needs the NVIDIA 550+ driver.
• No need to install the Visual C++ Redistributable: the CRT is statically embedded.

Installation (MSI)

1. Download Apollia-OS_<version>_x64_en-US.msi.
2. Double-click, follow the wizard. It names the product, the publisher (Apollia), the version and the dual MIT / Apache-2.0 licence.
3. The app shows up in the Start menu.

The release pipeline signs the installers with an Authenticode certificate only when it holds one. On an unsigned installer, SmartScreen warns about an unrecognised publisher: choose More info, then Run anyway.

The Windows firewall will ask you to allow apollia-os.exe, the llama-server.exe inference engine and apollia-runner-*.exe (speech-to-text) on first launch: allow them on private networks (these components talk to the daemon over loopback 127.0.0.1).

The installer downloads the WebView2 runtime if your machine does not already have it, so the install step needs a network connection on a machine that has never run a WebView2 application. Current Windows 11 ships it preinstalled.

Installation (.exe)

The NSIS installer asks who the installation is for. Install for all users puts it in C:\Program Files\Apollia OS\, the same place as the .msi. Install for me only needs no administrator right and puts it under your user profile instead. Note the destination the wizard shows you: the verification below reads from it.

Closing the app

The window close button quits Apollia on Windows, the way it does for any other Windows application. The runtime stops with it, and so do the two background processes it owns: the llama-server.exe inference engine and the apollia-runner-*.exe speech-to-text runner. macOS behaves differently, where closing a window is not the same gesture as quitting: there the app stays resident behind the menu-bar icon.

Quitting always stops the inference engine, whichever surface you quit from: the close button, the tray menu, or the in-app menu. Nothing is left holding video memory or a loopback port afterwards. To confirm, from PowerShell after quitting:

Get-Process apollia-os, llama-server, apollia-runner-vulkan, apollia-runner-cpu -ErrorAction SilentlyContinue

It should print nothing.

You should never see a terminal window belonging to Apollia. The background processes are console programs, and they are started with the Windows CREATE_NO_WINDOW flag so none of them opens a console of its own. If a terminal window does appear, that is a defect worth reporting rather than something to close by hand.

Verification

From PowerShell. The path below is the all-users one; on a per-user .exe install, replace it with the destination the wizard showed you.

& "C:\Program Files\Apollia OS\apollia-os.exe" --version
& "C:\Program Files\Apollia OS\apollia-os.exe" doctor

doctor runs eight checks: the data directory, the configuration file, the two databases, the models directory, Python, the sandbox posture and the runtime socket. It does not detect your GPU. The command that reports the detected hardware is a separate one, and it needs the daemon running:

& "C:\Program Files\Apollia OS\apollia-os.exe" model hardware --json

It answers with the total RAM, the CPU and the detected accelerator. Ask for --json: the plain-text rendering of that command prints nothing today.

What Windows does not confine

On Linux, a tool call runs inside namespaces with resource limits. On Windows there is no confinement at all: no namespaces, and no resource limits either, because the Unix mechanism has no Windows equivalent and the function that applies it does nothing on this platform. A tool an agent runs has the same rights over your machine as the application itself.

That does not make Windows unusable, and it does change what you should delegate. The permission rules and the approval prompts still apply, and they are the only barrier here, so treat an "always allow" on Windows as a wider grant than the same choice on Linux.

One practical consequence: bash_executor needs a POSIX shell on PATH, from Git Bash, WSL or MSYS2, and fails without one.

GPU acceleration

Two accelerations, and Windows answers them differently.

Local LLM inference already runs on the GPU. The installers embed a Vulkan build of the llama-server engine; the apollia-os-windows-x86-vulkan.zip command-line bundle carries the Vulkan one. Vulkan drives NVIDIA, AMD and Intel cards alike. Nothing to install and nothing to set: with a working driver the engine uses the card. The -cpu bundle is the one to take on a machine with no graphics driver.

An NVIDIA card can go further. There is one installer per system, and no separate one for NVIDIA cards. When the application detects such a card, its AI screen offers the accelerated engine, Speed up with your NVIDIA card: one click downloads it with the libraries it needs, after its signature is verified, and the next model load uses it.

Dictation runs on the GPU too. Speech to text runs in the apollia-runner sidecar, built on whisper, and the installer ships two of them: apollia-runner-vulkan.exe, which whisper drives through Vulkan, and apollia-runner-cpu.exe. The runtime picks the Vulkan one for an AMD or Intel card, and for an NVIDIA card without CUDA 12; it falls back to the processor only when no Vulkan loader is present. The apollia-os-windows-x86-vulkan.zip command-line bundle carries the Vulkan runner as well.

What is different on Windows

Windows is a supported platform, but three points set it apart from the other two and are worth knowing before you hand a task to an agent.

No tool confinement. On Linux, a command started by an agent runs in isolated namespaces and under resource limits; on macOS, under resource limits. On Windows, neither: a command started by an agent runs with exactly your rights, on your files, with no memory or CPU time cap. The practical consequence: on Windows, only run agents whose code you have read, and keep manual approval enabled in the chat.

The shell tool requires a POSIX shell. bash_executor looks for an sh in your PATH. Without Git Bash, WSL or MSYS2 installed, any agent that uses that tool fails. The other tools, files, web and Python, work normally.

A named pipe instead of a Unix socket. Elsewhere the command line reaches the runtime through a Unix socket that the filesystem protects at 0600. There is no such socket on Windows, so the runtime serves a named pipe, \\.\pipe\apollia-runtime-<user>, and the command line opens it. Two consequences for you. The global --socket option is accepted and ignored: the pipe name is derived from your USERNAME, not from a path you choose. And the pipe is created with a default security descriptor, weaker than the socket's 0600, so what protects the runtime here is the bearer token in %USERPROFILE%\.apollia\api-token. Treat that file as a password: any account that can read it can drive your runtime.

Update

See Update Apollia.

Uninstall

Settings > Apps > Apollia OS > Uninstall, for the .msi and the .exe alike. User data:

Remove-Item -Recurse "$env:USERPROFILE\.apollia"