Skip to content

Platforms and executor support

Choose an executor for your platform, then check its prerequisites. Host tasks are a quick starting point on Linux; choose a VM when you need Linux userspace or a separate kernel. VMs on Apple Silicon use Linux aarch64 userspace.

After installation, check that the controls you need are active. The executor observations in each Run Bundle identify that run's effective boundaries.

Mechanisms and prerequisites today

Prepare the environment using these prerequisites, then validate the features you need with a small task and its Bundle.

Platform host container VM
Linux x86_64 FUSE for staging; usable user namespaces and Landlock for sandbox Native crun/runc and a matching static Linux binary /dev/kvm; musl release embeds the guest kernel
Linux arm64 Source paths exist; check build and test evidence first A matching linux/arm64 injected binary Host/build KVM support and a matching rootfs; do not infer support from the x86_64 wheel
Apple Silicon macOS macFUSE 5.4.0+ FSKit for staging; Seatbelt for files and network Native OCI Linux-container path unsupported HVF, Hypervisor signing and a Linux arm64 rootfs/image; no macFUSE needed
Windows / Intel macOS No current release support promise No promise No promise

The container column describes the native Job OCI executor (--executor container). In the Docker-style container commands, pvisor run uses independent Linux VMs, with implementation paths on Linux x86_64/KVM and Apple Silicon/HVF. It does not use the OCI runtime in this column or implement the Docker Engine API.

Current wheels target Linux x86_64 and macOS arm64. Source branches do not establish released artifacts or maturity. Check pvisor --version, run a small task, and inspect its Bundle. See installation.

Single-node daemon

The sandbox API requires Linux x86_64/KVM and delegated cgroup v2; macOS HVF and Linux arm64 are outside this API's supported scope. prepare-image publishes templates in the shared compressed image pool, with the Rust guest engine as the default for new templates. Official OpenSandbox/E2B SDK core flows have real-VM validation; full protocol conformance and production density are not guaranteed. See the daemon guide for deployment.

The same executable's cache and memory-pool components support Linux and Apple Silicon macOS without enabling the sandbox API. macOS offers compression or explicit physical-pages sharing; see memory optimization for prerequisites and failure boundaries.

Verify capability evidence

  • Check safety.filesystem_read_non_bypassable and filesystem_write_non_bypassable separately, not just whether changes are staged.
  • Check safety.network_non_bypassable and the driver; selective Linux proxies and container host networking do not become mandatory boundaries just because of the executor name.
  • Container cannot satisfy --safe controls and rejects --safe. All current executors reject --strict due to the subprocess gap.
  • Inspect resources.effective and limitations. macOS does not enforce RLIMIT_AS; ordinary host rlimits are not total process-tree budgets.

Evidence entry points: .github/workflows/ci.yml, docs/src/zh/cases/06-stage-apply.md, crates/pvisor/tests/, and tests/test_vm_*. Untested cells receive no maturity claim. See boundaries.