Microsoft Open-Sources NVX: An Ultra-Light Micro-VM Sandbox for Agentic Workloads
Developed by MSR and Azure Research, NVX is an OpenVMM-based micro-VM sandbox offering hardware-enforced isolation for untrusted agent workloads. Here is how it
microsoft/nvx, an open-source micro-VM sandbox jointly developed by the MSR Systems Research Group and Azure Research - Systems, is available under the MIT License and actively maintained on GitHub. Built on top of OpenVMM (Microsoft's open-source virtual machine monitor), NVX runs a Linux guest kernel to execute untrusted AI agent workloads inside a hardware-enforced isolation boundary.

Image source: Coolcaesar 쨌 CC BY-SA 4.0 쨌 https://creativecommons.org/licenses/by-sa/4.0 쨌 https://commons.wikimedia.org/wiki/File:Building92microsoft.jpg
As AI coding agents and autonomous workflows increasingly execute arbitrary shell commands and generated code, isolating those runtime environments has become a primary engineering requirement. Standard containers share the host operating system kernel, meaning a kernel exploit can compromise the underlying host. Traditional virtual machines provide strong hardware isolation, but their boot latency and memory overhead make them heavy for short-lived, repetitive agent tasks. Drawing on research results from the Nanvix system, NVX strips out unnecessary components to deliver a lightweight micro-VM architecture tailored for isolated agent execution.
Architecture and Supported Guest and Hypervisor Environments
At the core of NVX is nvx-rs, a Rust crate and command-line interface that implements the OpenVMM control protocol along with launch, policy lowering, and supervision mechanics for micro-VMs. The repository tracks pinned versions of Linux, OpenVMM, and the Rust guest toolchain, alongside patches, guest source files, build scripts, and benchmark documentation.
- Supported Guest Userlands: NVX supports Alpine, Ubuntu, and Azure Linux userlands running on its own Linux kernel. Each guest image ships ten static guest binaries built from
nvx-rsat common/sbin/nvx-*paths to manage the single-workload guest runtime. - Cross-Platform Hypervisor Support: On Linux, NVX runs on standard KVM (
/dev/kvm) as well as Microsoft's MSHV (/dev/mshv). On Windows, it runs atop the Windows Hypervisor Platform (WHP).
Alongside the nvx-rs CLI, the repository documentation also covers scripts/nvx.py and direct OpenVMM integration paths for workflows that invoke OpenVMM without nvx-rs.
Quick Start and Execution Workflow
Developers do not need to compile OpenVMM or the guest kernel from source just to try NVX. Building the nvx-rs CLI automatically downloads a prebuilt NVX package matching the host platform on its first run. The host machine requires Git, Rust (via rustup and cargo), curl, and tar.
1. Linux / KVM
On a Linux host with read/write access configured for /dev/kvm, clone the repository, build the CLI, and launch a micro-VM:
git clone https://github.com/microsoft/nvx.git && cd nvx
cargo build --release --locked --features cli \
--manifest-path nvx-rs/Cargo.toml --target-dir nvx-rs/target
nvx-rs/target/release/nvx-rs run
2. Linux / MSHV and Windows / WHP
On Linux systems configured with read/write access to /dev/mshv, pass the --hypervisor mshv flag when running:
nvx-rs/target/release/nvx-rs run --hypervisor mshv
On Windows, enable Windows Hypervisor Platform first, then build and launch from PowerShell:
git clone https://github.com/microsoft/nvx.git; Set-Location nvx
cargo build --release --locked --features cli `
--manifest-path nvx-rs/Cargo.toml --target-dir nvx-rs/target
nvx-rs\target\release\nvx-rs.exe run
The first nvx-rs run fetches the prebuilt release artifact corresponding to the cloned revision (or its nearest released first-parent ancestor). After pulling newer commits, nvx-rs download updates the local artifacts, while users without a Rust toolchain can extract a development release archive from GitHub Releases and run its bundled bin/nvx-rs binary directly.
A successful boot prints NVX-GUEST-BOOT-OK: alpine and opens a root shell inside the guest. To boot into the Ubuntu or Azure Linux userland instead, append --guest ubuntu or --guest azurelinux to nvx-rs run. Inside the guest, running /sbin/nvx-exit 0 performs a clean shutdown.
Verified Prerequisites and Limitations
Created on August 27, 2026, the microsoft/nvx repository continues active development with v0.1.0-dev development release tags published as of November 13, 2026. Teams evaluating NVX should keep several verified prerequisites and boundaries in mind:
- Submodule Initialization for Full Source Builds: A standard
git cloneomits theopenvmmsubmodule. Anyone building OpenVMM and the guest components directly from source rather than relying on prebuilt release packages must initialize theopenvmmsubmodule first. - Host Hypervisor Permissions: Hardware virtualization support is mandatory. Linux hosts require read/write permissions on
/dev/kvmor/dev/mshv, and Windows hosts must have Windows Hypervisor Platform enabled. - Unverified Benchmarks and Stripped Guest Scope: While
doc/benchmarks.mdprovides instructions for running benchmarks, specific boot latency or memory footprint numbers are not reported in the main README. Because NVX strips out unnecessary components for minimal overhead, workloads requiring full-featured system daemons or specialized kernel modules should be tested individually for compatibility.
Sources
- GitHub (
microsoft/nvx): https://github.com/microsoft/nvx