FTL, a userspace OS for clouds, reaches v0.1.0 with async Rust support
FTL, an experimental operating system that moves most of the OS into a shared library running in userspace, published v0.1.0 with support for async Rust apps, Linux threads and epoll, according to its project site. The project calls itself very alpha quality.
FTL is an operating system for cloud servers, and it is not finished. Its project page, updated on 3 October, lists two milestones. The first is a Linux HTTP server running on FTL in v0.0.1 in September 2026. The second is async Rust app support, including Linux threads and epoll, in v0.1.0 in October 2026. That is the state of the release. No version number beyond v0.1.0 is given.
The design is the interesting part. FTL keeps a small kernel. It pushes the rest of the OS into a library that each container loads: processes, virtual file system, TCP/IP. The kernel then exposes a hypervisor-like interface to that library. Linux system calls get implemented in userspace rather than in the kernel. The site's own diagram shows Linux processes calling into a userspace OS. That OS talks to a minimal kernel handling vCPU, memory and drivers. It is a clean picture, and the project is candid that the picture is ahead of the code.
What the kernel keeps, and what it gives away
The project claims this gives stronger isolation between containers than monolithic kernels achieve, without hardware-assisted virtualisation. It also says you can update the userspace OS without restarting the machine. Each container can run a different OS or application kernel version. It also says Linux binaries are compatible. The Rust-based HTTP server serving the FTL website is a Linux application running on FTL. Specialised applications can run without POSIX abstractions.
In a blog post introducing the project, also dated 3 October, the author writes that he began with a microkernel design. He changed the architecture several times. The goal is to let developers try experimental ideas such as io_uring, observability hooks or a custom TCP stack without sudo or kernel programming. He names gVisor's Sentry application kernel as a comparable idea. He says FTL is still "very alpha quality" but "started working." He also concedes the obvious objection: Linux is mature and battle-tested, and FTL is neither.
The kernel focuses on providing a minimal interface to it, the blog post says of the userspace OS library.
That is a familiar promise. Microkernels, unikernels and sandboxed container runtimes have all argued that a smaller trusted core is worth the cost. What FTL adds is the claim that the cost is small: a hybrid kernel that keeps monolithic performance while giving the flexibility of a microkernel. No benchmark numbers support that claim, and the project page offers none. The performance side of the argument is currently an assertion rather than a measurement.
The release sits inside a busy week for small systems
FTL was not the only operating system news in the last three days. IINA, the macOS media player, shipped version 1.5.0 on 3 October. It is its biggest interface redesign since launch. It includes a new Layout sidebar, a rewritten Settings window built with Swift result builders, and plugin discovery inside the app. On the Amiga side, a developer published ODFileSystem, also on 3 October. It is an open-source read-only optical disc filesystem for AmigaOS 3 and AmigaOS 4. It supports ISO 9660, Rock Ridge, Joliet, UDF, HFS, HFS+, High Sierra and CDDA audio tracks.
ODFileSystem's README makes a measurable claim rather than a general one. In CDBench, which publishes results for all eight filesystems, its 68020 build is fastest in every metric tested. Sequential reads reach 205,197 KiB/s. Small-file open and read take 319 microseconds, more than three times faster than the next best. The 68000 build leads in every metric except random reads from the same offset. Those are the kind of numbers FTL does not yet publish.
Both projects share a trait worth noting: very small teams maintain them. IINA says so directly in its release post, noting that development can move slowly because of limited resources. FTL's author describes a year of work by one person. Racks of GPU instances behind the CS:GO dataset work at Vortech, an unrelated three-person team, serve as a reminder of how thin these operations usually are.
Agents, tokens and the tooling layer
The same week produced a cluster of developer tools aimed at people running AI agents. Several of them touch the operating-system layer. AgentSight, published on GitHub on 3 October, is a local-first observability tool. It uses eBPF and TLS traffic tracing to show what an agent actually does on a machine: processes spawned, files changed, services contacted. Its pitch is that it needs no SDK, no proxy and no vendor integration. So it works on closed-source CLIs such as Claude Code, Codex and Gemini CLI.
A separate audit published the same day by the RugSnare project claims 218 silent contract changes across the most-installed MCP servers on npm. It says none were announced in a changelog. Among its findings: 75 tools removed between minor versions of Microsoft's @azure-devops/mcp, and all 28 schemas changed in chrome-devtools-mcp. The project cites two third-party datasets that it says show a similar pattern. One reports that 1,419 of 21,738 packages changed their tool definitions in four weeks. These are the audit author's own numbers and have not been independently checked here.
Elsewhere, Meta AI Research said on 3 October that it had worked with mathematicians using its Muse Spark models in Thinking Mode through the ordinary chat interface. There was no custom research scaffold. It was sharing six papers, five of which present answers to previously open questions. Meta notes that other teams independently announced solutions to some of the same problems and that the papers acknowledge them. Aleph Alpha released Kolibri on the same day. It is an English-German mixture-of-experts model with 78B total parameters and 3B active, under Apache 2.0.
Why a cloud OS matters beyond the demo
The context for FTL is a market where the container has become the unit of deployment. The kernel underneath it has become a shared liability. A container escape is a kernel bug away, and patching the kernel means rebooting the host. FTL's answer is to make the OS a library per container. A fix then ships like an application update, and each tenant can run a different version. Whether that holds up under real workloads, and what it costs in system call overhead, is exactly what the project has not yet shown.
There is also the question of who would run it. The site points to Google Compute Engine as a place to try it. The blog post frames the target as cloud computing rather than hobby projects. That is an ambitious frame for a v0.1.0 release with a Linux HTTP server as its headline demonstration. The gap between a working demo and a production cloud OS is measured in years. The project's own description of its quality sets expectations accordingly.
For now the honest summary is this: FTL is a serious idea with a working prototype, published openly. Its design document explains the trade-offs and offers no performance data to settle them. The code and the write-up are public. The benchmarks are the next thing to watch.
Sources
12- 01FTL: A new operating system for cloudsEN
- 02Introducing FTL: A new operating system for cloudsEN
- 03IINA (the media player) 1.5.0 is ReleasedEN
- 04A modern open-source Amiga CDFileSystem replacementEN
- 05AgentSight: System-wide AI agent profiling and monitoring with eBPFEN
- 06We diffed all 66 release pairs of the official MCP servers, 140 silent changesEN
- 07Solving Open Research Problems TogetherEN
- 08Kolibri: A Sovereign Open-Weight ModelEN
- 09Open 10-player CS:GO dataset for multiplayer world modelsEN
- 10Cloudflare opens more Enterprise-only features to all customersEN
- 11California Attorney General Serves Investigative Subpoena on OpenAIEN
- 12OpenAI says its review into hacks, including on Australian government sites, is costing $500,000 a dayEN
All figures and quotations in this text come from the sources listed below.
Content prepared by the editorial team with AI assistance.
Comments
0- No comments yet — be the first.