CentOS Founder Builds Linux-Compatible Rust Kernel with AI in Just 11 Days

CentOS and Rocky Linux founder Gregory Kurtzer unveiled Ginux, an experimental Linux-compatible kernel written mainly in Rust that was developed with AI assistance in 11 days, according to his technical report. A single persistent OpenAI Codex task using the GPT-5.6 Sol model generated approximately 139,000 lines of retained implementation and test code, producing a kernel that covers 375 Linux 6.12 x86-64 system calls — 82 fully implemented, 245 partially implemented and 48 deliberately denied — and can boot a Rocky Linux 10 userspace through systemd's multi-user.target. Kurtzer, who calls the method Human-Directed Agentic Engineering, released Ginux under the Apache 2.0 license on GitHub but states it remains a research prototype, not production-ready, with incomplete Linux ABI coverage and limited hardware support.

Gregory Kurtzer, founder of CentOS and Rocky Linux, has unveiled Ginux https://x.com/gmkurtzer/status/2114958282481910115 , an experimental operating system kernel written mainly in Rust and developed with AI assistance in 11 days. Unlike Linux distributions, Ginux uses an independent kernel designed to run existing Linux userspace applications without relying on the Linux kernel. The project was developed as an experiment in what Kurtzer calls Human-Directed Agentic Engineering, a methodology where a human developer defines architecture, requirements, constraints, and acceptance criteria, while an AI coding agent handles implementation, testing, debugging, and revisions. According to Kurtzer’s technical report, a single persistent OpenAI Codex task using the GPT-5.6 Sol model generated approximately 139,000 lines of retained implementation and test code over the 11-day development period. Despite the short timeframe, Ginux includes its own memory management, task scheduler, process model, virtual filesystem, networking stack, device support, and Linux ABI compatibility layer. It targets x86-64 systems and boots directly through UEFI without requiring an underlying Linux kernel. One notable achievement is the ability to execute selected unmodified Alpine Linux applications using musl and BusyBox, along with Rocky Linux programs using glibc. The prototype can also boot a Rocky Linux 10 userspace running systemd through the multi-user.target stage. Regarding Linux compatibility, Ginux covers 375 system calls from the Linux 6.12 x86-64 interface. Of these, 82 are fully implemented, 245 partially implemented, and 48 deliberately denied. Architecturally, Ginux organizes workloads into isolated environments called domains. Each domain maintains its own processes, address spaces, filesystems, mounts, capabilities, and lifecycle. The system can also retrieve and assemble OCI container images, using them to provide Linux userspace environments while Ginux itself handles execution and isolation. The kernel, embedded administration controller, and essential boot components are packaged into a single GINUX.EFI image. The controller runs in user space, not kernel space, and manages configuration, devices, networking, and domain operations. For now, development and testing primarily target virtual machines using QEMU/KVM, with an x86-64 Q35 configuration, OVMF firmware, Intel IOMMU, and virtio devices. The repository also documents execution through QEMU’s emulation mode on Apple Silicon. However, Kurtzer makes clear that Ginux remains a research prototype, not a production-ready operating system. Linux ABI coverage is incomplete, hardware support is limited, and areas including networking, filesystem functionality, resource enforcement, and security still require substantial development. According to Kurtzer, the main purpose of the experiment is to evaluate whether AI coding agents, guided by experienced developers and supported by executable tests, can rapidly implement complex systems software. Ginux is available as open-source software under the Apache 2.0 license on GitHub https://github.com/gmkurtzer/ginux . Those interested in the development methodology and technical findings can also consult Kurtzer’s research report https://zenodo.org/records/23682548 .