Blog · · 5 min read

What is Hermes Desktop? A plain explanation for business owners

Hermes Desktop is the app you open on your computer to work with an AI agent called Hermes Agent. Nous Research makes both, and publishes them as open-source software under the MIT license[1][2]. The desktop app runs on macOS with Apple Silicon, on Windows and on Linux[1]. Kodros installs it, pins a version so it does not move under you, and pairs it to the server in your office. We do not make Hermes, and it is not Anthropic's software.

By Tanner Osterkamp

What an AI agent is

A chatbot answers the question you typed and waits. An agent also takes steps. It reads files, calls tools, drafts a document and works down a list you gave it. Hermes Agent is that second kind. It keeps memory between sessions, and it uses skills, which are saved instructions for a kind of job[2]. The memory and the skills are where the useful work comes from. They are also why you want a person in the loop before anything is sent. Our post on AI agent approvals is the set of limits we actually turn on.

What the desktop app does

Hermes has several front ends, and they all talk to the same agent. The desktop app is the one built for someone who does not want to live in a terminal. The Hermes documentation says it gives you[1]:

  • A chat window that shows the agent's work as it happens, and then the final paragraph.
  • A file browser, so you can see which files it read and which it wrote.
  • A preview pane for web pages, files and results, sitting next to the chat.
  • Voice, if you would rather talk and listen than type the whole exchange.
  • Settings pages for providers, models, tools and connected services, so a model change is a screen and not a config file.
  • The same conversation history as the other Hermes front ends. A session you start in the app can be picked up elsewhere.

Hermes works with hosted model providers and with models you run yourself[2][3]. The model is the program that writes the answers. The choice of model is also the choice of where your questions and files get read. Pick that on purpose, before the first real job, not after a staff member has already pasted a contract into a hosted box.

Where your data goes

When a model runs on your own hardware, the questions and files are read there. The Hermes documentation describes running models entirely on your own machine, with no account and no API key[3]. When you pick a hosted provider, the prompts you send are processed by that provider, under their terms, on their machines.

In a Kodros setup the local model runs on your server. A job may leave that server for four reasons, and only those: the local model cannot do the job well enough, it is too busy, it is down, or a named person says to send it out. You choose which of those reasons are allowed, and you can allow none of them. Connectors such as email still reach outside apps, because that is what a mail connector is for. How it works lists what can still leave the building even when the model stays home.

How the app and your server fit together

Hermes can run on one machine and be used from another. The documentation describes connecting the desktop app to a Hermes instance on a remote machine[1]. That is the shape we install. The server in the office holds the shared files and records. Hermes Desktop on your computer is how you hand it work and watch it. The Kodros iPhone app is for the short questions and for approvals when you are not at the desk. Day 5 of the one-week install is when your devices get paired to that server, after the restore test has already passed.

Why we pin the version

Hermes is updated often, and a new version can change how a tool behaves on a Tuesday morning when nobody asked for a change. Pinning means we lock a version in place. On day 2 of the one-week install we install the software, pin that version, and write what was installed into the handover record. A later update is a decision, with a reason, not something that arrives because a background updater felt like it.

What Kodros adds

The software is open source. You can install it yourself, and Nous Research publishes the documentation and the code[2]. What you are paying Kodros for is the work around the app:

  1. A server in your office that you own, with your files and records on it.

  2. Backups, and a restore test that you watch before you sign.

  3. Hermes Desktop installed, pinned and paired to that server.

  4. A named person who approves sends, payments, posts and access changes.

  5. A handover record: what was installed, who owns each account, and where the backup lives.

We do not make Hermes; Nous Research does. If a vendor tells you Hermes Desktop is an Anthropic product, they have the wrong app.

Is it a fit for you?

Hermes Desktop fits when the work is documents, files, email and the repeated tasks an assistant could prepare for you to review. It does not remove the need to read the output. The model can be wrong, and a fluent wrong answer still looks finished. Start with one small job and a named person who checks it. If the job is a single chat box and you do not have files to gather, a rented seat may be the smaller project. The comparison is in rent or own.

Questions

What is Hermes Desktop?

A native app for macOS, Windows and Linux for working with Hermes Agent, an AI agent from Nous Research. Chat, tool activity, files and previews sit in one window[1].

Does Kodros make Hermes Desktop?

No. Nous Research makes it and publishes the source under the MIT license[2]. Kodros installs a pinned version and pairs it to your server. We are not Nous, and the app is not Anthropic software.

Is Hermes Desktop free?

The software is MIT-licensed open source, so the license itself does not charge you per seat[2]. A hosted model provider can still bill for use. Running a model on your own hardware adds no per-question fee, and it does require the hardware and the power to run it.

Which AI models does Hermes work with?

The Hermes documentation lists Nous Portal, OpenRouter, OpenAI, Anthropic, Google and any OpenAI-compatible endpoint, plus local models[3]. Choose by where you want the prompts read, and by whether that model can do the jobs you wrote down.

Why does Kodros pin the version?

So the software does not change on its own. Pinning locks a version in place. An update happens when someone decides to update, and the handover record says what is installed now.

Sources

  1. Nous Research, Hermes Desktop, Hermes Agent documentation. hermes-agent.nousresearch.com
  2. Nous Research, hermes-agent, source repository and MIT license. github.com
  3. Nous Research, Hermes Agent documentation index and Local Models. hermes-agent.nousresearch.com

What to do next

Kodros puts private AI on hardware you own: the model, the app and the compute, in your building. Every send, post and payment waits for a person. If you leave, you keep the hardware, the files and the memory.

More to read