Image: Herald OS

An independent developer who goes by Luke The Dev has released Herald OS, a free, open-source desktop built around an AI agent rather than around folders and menus. You talk or type to the agent, and it opens apps, finds and sorts files, runs routines and builds small pieces of software on the machine.

The agent is Hermes Agent, the open-source agent published by Nous Research. Herald OS is not a Nous Research product. The developer says so in the announcement, and the project's files state that it is not affiliated with, sponsored by or endorsed by Nous Research.

I built an operating system where the AI agent isn’t another app. It’s part of the OS.

Today, Herald OS is open source and you can install it yourself 🚀

It’s built around Hermes Agent by Nous Research.

— Luke The Dev (@iamlukethedev) October 7, 2026

The announcement lists a lot of features: a "Hey Hermes" voice wake word, asking about any part of the screen, plain-English explanations when an app crashes, one-click installs for coding tools like Claude Code and Codex, and themes that can be generated from a photo or a text description. These are the developer's claims. We haven't installed Herald OS, and nobody has independently tested how well the agent actually performs these tasks.

"Operating system" means two different things here. On a PC, Herald OS Linux is a full system that installs from a USB stick: Fedora handles the kernel, drivers and packages, and Herald draws everything you see on top. On a Mac, it is an app that runs fullscreen over macOS. There is also a package for Arch Linux, an add-on for the Omarchy desktop, and a virtual machine that Apple Silicon Mac owners can boot with one command.

The code shows that Herald OS does not modify Hermes. It runs a standard Hermes install and adds its computer controls as a regular Hermes plugin. That keeps the project light, but you still have to bring your own brain. Hermes needs a model provider, either a Nous Portal account or an API key from a service such as OpenRouter, OpenAI or Anthropic, or a local model. Depending on what you choose, using the agent can cost money.

The developer says basic OS commands don't burn model tokens and that Hermes only gets involved when you ask it to think or do a task. That matters for cost, because an agent that sends every click to a paid model gets expensive fast.

Giving an agent access to your whole computer is the obvious risk, and the project's design documents say how it handles that. Each tool the agent can use is sorted into one of four tiers. "Read" tools, like checking memory use or searching files, run straight away. "Act" tools, like opening an app or a link, also run without asking but show a notification. Changes such as moving or renaming files need your approval. Destructive actions, such as killing a process or trashing files, ask every single time, and that tier can't be set to auto-allow.

Some locations are blocked outright before any prompt appears, including SSH keys, password keychains, browser cookies, the agent's own credential files and core system folders. Every tool call is written to an audit log on the machine. Files are sent to the trash and never deleted with a hard rm.

There are two caveats in the fine print. The approval prompts use Hermes's own approval system, so if a user turns approvals off or switches on Hermes's "yolo" mode, Herald honours that setting. And because "act" tools run without asking, the agent can open apps and web pages on its own. The project's security policy treats any way around these checks as a vulnerability, and it asks people to report such issues privately.

It is also a very young project. Its public history goes back only about three weeks, and all 62 commits come from one developer. The first release is labelled 0.1.0-alpha.1. The README warns of rough edges and tells users to keep backups of anything they let the agent touch. The Mac app isn't notarized by Apple yet, so macOS blocks it until you clear a security flag by hand. The Omarchy integration has been written against Omarchy's configuration but, by the developer's own account, has not been tried on an actual Omarchy machine.

The idea isn't new. Assistants like Claude Code and Codex already work on people's computers from a terminal or an app window. What makes Herald OS different is that the agent sits at the centre of the desktop itself, with permissions built into the system rather than added through a chat window.

Whether that is a better way to use a computer is something early testers will have to decide. For now, Herald OS is an MIT-licensed experiment that anyone can boot in a VM, and the code is public for anyone who wants to check how much control it hands to the agent.