App Sandbox is something special. It’s free and open-source. Windows 11 Home can run Windows 11, Ubuntu, and macOS VMs without enabling Hyper-V, and it comes with GPU acceleration. The coolest part is the headless API plus Python SDK, which lets you programmatically spin up multiple VMs and feed them to AI agents — one laptop can run several at once. 🔗 https://t.co/wfuNkqq63p https://t.co/XaqgQsC0av

jamesstringer90/appsandbox Source: https://github.com/jamesstringer90/appsandbox

App Sandbox

App Sandbox is a virtual machine app for Windows and macOS that’s focused on performance and ease of use. Windows features:

• Works on Windows 11 Home or Pro, without Hyper-V
• Windows 11 or Ubuntu 26.04 LTS VM Support
• Zero touch install
• Copy and Paste
• 2 Channel Audio
• GPU Acceleration via Paravirtualization (GPU-PV) with support for DirectX 12 (Windows only), OpenGL, Vulkan, CUDA, OpenCL
• GPU Hardware Video Decoder/Encoder support
• SSH via Hyper-V socket proxy (no network required)
• Snapshots
• Fixed 1080P60 display
• Host to client hot-key support
• Provision and boot with / without internet
• Supports running Claude Cowork and Docker inside the VM thanks to Nested Virtualization Support
• Headless mode: a scriptable local HTTP/JSON API + dependency-free Python SDK (asb.py) for creating and driving GPU-accelerated VMs programmatically

Mac features:

• macOS VM support
• Skips most of the install setup process
• Copy and Paste
• 2 Channel Audio
• GPU Acceleration via Paravirtualization with support for Metal
• SSH via virtio-vsock (no network required)
• Dynamic display sizing
• Provision and boot with / without internet
• Headless mode: the same scriptable local HTTP/JSON API + Python SDK (asb.py) as on Windows, for driving macOS VMs programmatically

Requirements: Windows 11 with an x64 Processor or macOS Tahoe (M Series)

About App Sandbox

App Sandbox creates and runs full desktop virtual machines — Windows 11, Ubuntu, and macOS. It is free and open-source (MIT), distributed as prebuilt binaries (EV-signed on Windows, Apple-Developer-signed on macOS) on its Releases (https://github.com/jamesstringer90/appsandbox/releases) page.

It is driven two ways: from a graphical UI, or programmatically through a headless daemon with a dependency-free Python SDK (asb.py). Either way you pick an OS, point it at an installer image, and App Sandbox provisions the disk, runs an unattended install, and boots the guest. You supply a Windows or Ubuntu ISO; macOS guests download their restore image automatically.

It runs on a Windows 11 (x64) PC or an Apple Silicon Mac, including laptops.

Runs on Windows 11 Home, without Hyper-V. On Windows, App Sandbox does not use Hyper-V or Hyper-V Manager; it creates and runs VMs through the Windows Host Compute System (HCS) and Host Compute Network (HCN) APIs, which require only the Virtual Machine Platform feature. Hyper-V is limited to Windows 11 Pro and Enterprise, while Virtual Machine Platform is available on Windows 11 Home, so a Windows 11 Pro license is not required. Virtual Machine Platform is the same Windows feature WSL2 uses.

What you can use it for:

• A full Windows or Ubuntu desktop VM — a complete graphical OS with GPU acceleration (OpenGL, Vulkan, CUDA) — for building and testing your own software on a clean install.
• Running a program in a VM, kept separate from your main OS and files.
• Developing and testing Windows drivers: a built-in Test Mode enables test-signing inside the guest, so the host’s Secure Boot and boot configuration are left unchanged.
• Running a computer-use AI agent (such as Claude’s computer use model) in its own VM. App Sandbox provides the desktop and the VM; the agent’s own model loop runs inside the guest. Several agents can run at once, each in its own VM, on a single laptop or desktop.
• Creating GPU-accelerated (GPU-PV) VMs, from the GUI or the headless API.
• Scripting the VM lifecycle from code — create, snapshot, branch, SSH into, and delete GPU-accelerated VMs — with the Python SDK.

Headless API. appsandbox.exe --headless (or sudo ... --headless on macOS) starts a single-owner daemon that hosts the same core as the GUI and exposes it as a Docker-style local HTTP/JSON API on 127.0.0.1, identical on both platforms. The stdlib-only Python SDK (asb.py) wraps it: create GPU-accelerated VMs (GPU-PV is a gpuMode create option), SSH in over an auto-deployed key (Windows, Ubuntu, or macOS guests), snapshot and branch them, open the live display, and run several at once — how many run concurrently is bounded by host CPU, RAM, and GPU memory, not by a fixed limit. The daemon is single-host — it drives VMs on the machine it runs on and does not provision cloud VMs; for CI, run it on a self-hosted runner. Provisioning a VM runs a full unattended install; snapshots, branches, and templates (Windows-only) start a new VM from a provisioned state instead of reinstalling. A VM persists until deleted. A snapshot is a disk-state checkpoint taken with the VM stopped, and a branch forks a writable disk from it to run divergent actions. The full API reference and runnable examples are in tools/headless-api/.

How it works: this repo also aims to be a working example of creating full desktop VMs programmatically with the Windows HCS/HCN APIs and Apple’s Virtualization.framework. On Windows, App Sandbox submits a hand-built HCS machine document to computecore.dll / computenetwork.dll — the HCS/HCN layer that also underlies WSL2 and Windows Sandbox — rather than going through Hyper-V Manager. GPU acceleration uses GPU paravirtualization (GPU-PV) to share the host’s installed GPU with the guest — the same WSL2 path, not a dedicated passthrough (no VFIO/IOMMU, no second GPU). Windows guests get DirectX 12, OpenGL, Vulkan, CUDA, and OpenCL; Ubuntu guests the same minus DirectX; macOS guests Metal. A custom IddCx indirect display driver and a virtual audio device carry the screen and sound, and guest↔host clipboard, audio, input, and SSH run over Hyper-V sockets. Guest disks are built by an in-repo tool with support for ext4, squashfs, qcow2, and VHDX. On macOS, App Sandbox uses Apple’s Virtualization.framework (VZVirtualMachine, VZMacOSInstaller) over virtio-vsock. On Windows, Linux (Ubuntu) guests reach the GPU through a custom DRM/KMS kernel module (asb_drm), Microsoft’s WSL2 dxgkrnl, and a custom Mesa build. The apps are native C / Objective-C with an HTML/JS UI (WebView2 on Windows, WKWebView on macOS).

Tips:

[Windows] Enable hotkeys or mute the VM audio: connect to the VM and right-click the connection title bar [Windows] Need a high performance remote desktop to remotely access your VM? Phaze (https://phaze.app) works well [Windows] You can check if the GPU-PV driver setup is working by running gpu-test.exe inside your App Sandbox Windows VM, gpu-test.exe will show a box with 6 rotating cubes, each using a different rendering engine (D3D9, D3D10, D3D11, D3D12, OpenGL and Vulkan). If one or more fail, they will not correctly show a rotating cube for that rendering API. gpu-test.zip (https://github.com/jamesstringer90/appsandbox/releases/download/v0.1.2/gpu-test.zip). Note: The rendering API succeeding means that the GPU-PV worked, but sometimes games or apps are coded in such a way that they will not detect the GPU-PV system correctly and still show an error.

License

AppSandbox is licensed under the MIT License — Copyright (c) 2026 James Stringer. It bundles third-party components that are not covered by the MIT license and retain their own terms — most notably the Microsoft WSL2 dxgkrnl GPU driver (GPL-2.0) and the embedded XZ decoder (0BSD). See THIRD-PARTY-NOTICES.md for the full list. AppSandbox’s own Linux DRM kernel module (tools/linux/asb_drm/) is dual-licensed MIT OR GPL-2.0 so it can resolve the kernel’s GPL-only symbols.

Acknowledgements

Beyond my own experience building Easy-GPU-PV (https://github.com/jamesstringerparsec/Easy-GPU-PV), I found NanaBox (https://github.com/M2Team/NanaBox) to be a really helpful resource for understanding HCS.

Author

James Stringer (https://www.linkedin.com/in/jamesstringerphotography/) — author of Easy-GPU-PV (https://github.com/jamesstringerparsec/Easy-GPU-PV).