Maclet puts its own Rust container runtime inside a Mac’s Linux guest
Sumant Sogikar built the isolation machinery after tiring of his development setup’s RAM use; Apple’s own container project offers another point of comparison.

The small number on Maclet’s README is a roughly 3.6 MB binary. Its author also wrote the container machinery inside the Linux guest, down to the calls that isolate a process and limit its memory. Saving RAM apparently led to a rather substantial side project.
Sumant Sogikar describes the motivation in his own account of building Maclet: dissatisfaction with Docker Desktop’s resource use, followed by a question about building a runtime without Docker Desktop, containerd or runc. Those familiar engines handle much of the work Maclet takes on itself.
Maclet’s architecture document puts the command-line tool and management daemon on macOS. The host handles images and cached layers, then Rust reaches Apple’s Virtualization Framework through a Swift interface. A minimal Linux guest supplies the kernel facilities that containers need. “Native macOS” describes the host software; your Linux program still has a Linux kernel underneath it.
Inside the VM, a static Rust guest agent takes the role of PID 1. It creates the cgroup hierarchy, sets up network namespaces and builds filesystem overlays before spawning the container’s process. The host has a dedicated virtual-console channel for commands, separate from the boot log channel. That means managing the guest does not first require discovering its network address.
The technical-constraints document states the boundary plainly: “All container isolation happens inside a real, lightweight Linux virtual machine.” macOS lacks the Linux namespace and cgroup interfaces this design uses. Copying a shipping container’s label onto a Mac process would leave some fairly large holes in the walls.
The README shows how much that guest agent owns. Namespaces separate what a process can see; cgroups enforce memory, CPU and process-count limits. OverlayFS gives each container a writable layer over its image. The image client resolves ARM64 manifests and verifies downloaded layers with SHA-256. Networking adds a Linux bridge, virtual Ethernet pairs and port forwarding. Maintaining those features means dealing with the same image and network complications that developers bring to larger runtimes.
Apple has its own container tool, written in Swift and optimized for Apple silicon, that also runs Linux containers using lightweight virtual machines. That provides useful context for Maclet’s premise: native Mac tooling and a Linux guest can coexist. The design choices inside and around those guests still need comparing on actual applications.
Sogikar’s boot and idle-memory figures are maker claims. The comparison table gives alternatives much larger numbers, but it does not replace a matched test of the same image on the same machine. Even an empty VM and a VM with a running database answer different questions about memory use.
A lean runtime could appeal to developers who need a few predictable services and are willing to inspect rough edges. The harder demonstration is an ordinary project with persistent data and awkward networking. If that project behaves correctly after stopping and restarting the VM, the savings become something a developer can use. Which everyday workload will supply that independent comparison?