The Hermes Agent — Tell It in Words and It Operates Your Computer for You

Hermes isn't an AI that just answers questions. It's an open-source agent that takes your orders and performs actual computer work. It takes its name from Hermes, the messenger god who carried the gods' messages in Greek mythology. You give instructions in plain words in a chat window, and it breaks them into a plan, manipulates tools, and produces the deliverable. This post is a writeup of installing the agent myself and using it in a local environment.

How is this different from existing AI

The difference between conversational AI and Hermes is easy to see with an example.

Ask for a Jeju flight timetable and ChatGPT-type AI finds the information and hands it back as text. With Hermes, tell it your Jeju business trip schedule and it goes and checks your calendar, searches for the lowest fare, and prepares the screens it needs. In other words, the difference between a guide that gives you information and a secretary that does the work.

Four things at the core

First, it operates tools directly. Not just prose — it runs and handles software like a web browser, documents, and a calculator.

Second, it makes its own plan. Give it something multi-step, like a summary email to send to your team, and it breaks it into a sequence — open the file, summarize the contents, write the email, send it.

Third, it's open-source, so you can keep it on your own computer. It's an MIT-licensed project published by Nous Research, and you can find it in the NousResearch/hermes-agent repo on GitHub. That means running it locally without sending personal or work data to an external service.

Fourth, you get to pick which model to attach. Connect a local model like Ollama and you can learn the structure or run light tasks without paying per token. Attach paid API models like Claude or GPT and it can take on work-grade tasks. That said, local models are closer to personal study and learning — for real work like the examples below, an API model is the realistic choice.

The "low canned refusals" point comes up often too. Especially in a local-model setup, the refusals you see all the time on commercial services are comparatively rare. That said, it varies with the character of the model you're using.

How you actually use it day to day

The examples below assume attaching a paid API model like Claude or GPT. With a local model alone, the polish doesn't get to where you'd want.

Selling used stuff. When you want to sell an unused bike on a secondhand site, you can instruct it to check market rates and write the listing.

Hospital appointments and prep. Ahead of a big screening appointment, tell it to check available times and the things you need to prepare, and have it organize them into a table.

Travel prep. Before departure, instruct it to bundle exchange rate info, the documents you need, and the basic itinerary into a one-page list.

Photo organizing. It can also group the photos piled up on your phone by year and make a backup folder.

Put simply, the most accurate description is: you give instructions in words, and it's a worker that turns them into computer manipulation.

How it felt to use

We installed Hermes directly in a Linux environment. Installation went through the official script, and it offers two forms: a desktop app and a terminal interface. Type the hermes command in the terminal and a conversation starts; the desktop app can be configured to launch under a separate name.

Wiring up a local model was easy too. Prepare the model in Ollama, then specify the model and endpoint in the Hermes config. In our environment, a 4GB-class local model on the GPU gave responses in around a second, and bigger models changed speed noticeably depending on hardware specs.

That said, this experience was only about learning the structure and trying light tasks with a local model. For real work, moving toward an API model is the realistic path.

There are a few things to watch in real use.

One is speed. If the local model is slow, you can feel like it "won't answer." We actually hit a case where a 9GB-class model ran out of VRAM and part of it spilled onto the CPU, making responses much slower. Dropping to a smaller model or shortening the context length makes a clear difference.

Second is configuration. Hermes is built to use several reference models and auxiliary features together, and if an unused external model connection is left on, responses can stall or lag. You need the habit of checking which model is actually being used when you start.

Third is the input environment. When typing Korean in the terminal, the enter-key handling can differ depending on your input method settings. That's not a Hermes problem in itself — it's the terminal-and-IMe combination — but it can confuse someone setting things up for the first time.

With those three in mind, Hermes reads as a tool faithful to its definition: "a worker that operates the computer when told to." That doesn't mean the automation is perfect all the way through. For important work, verify the results, and tune speed and configuration to your environment.

Verdict

Hermes sits on the extension line of conversational AI, except the output leads to task performance rather than text. It's open-source and you can attach a local model for free, which appeals to users who care about cost and privacy.

Keep your expectations realistic. It has the framework for planning and manipulating tools, but how well it does depends on the connected model and environment. "A worker that does it for you when you tell it" is an accurate description. How quick-handed that worker is, that's something you judge by attaching one yourself.