AI Talk - Hermes Agent

If you've spent any time around local AI this year, you've run into Hermes Agent. It's open source, it runs on hardware you control, and it can use whatever model you point it at, including one running on your own box. For someone like me who's getting into running models locally, that's most of the appeal right there.

Then, in the span of three weeks, Meta shipped Muse and OpenAI shipped Dots. Both are always-on agents that keep working after you close the app, and both come from companies with a few hundred million more users than Nous Research. So the obvious question: does a self-hosted agent still matter when the big platforms hand you one for the price of a subscription you might already have?

I think it does, but not for everyone. Here's my attempt at sorting out who.

What Hermes Agent actually is

Hermes Agent is an open-source agent from Nous Research, the lab behind the Hermes fine-tunes of Llama. It was released under the MIT license in February, and as of this writing the GitHub repo has about 250K stars. Earlier this month Nous raised $90M to build a business version, and the Wall Street Journal put the company at a $1.5 billion valuation.

The simplest way I can describe it: it's the harness, not the model. Hermes is the loop that takes your request, decides which tools to call, runs them, remembers what happened, and keeps going. The thinking is done by whatever model you plug in, whether that's Nous's own Portal, OpenRouter, OpenAI, Anthropic or your own endpoint. Switching is a single hermes model command.

Interesting things:

• Persistent memory. It keeps notes on you and your projects across restarts, and it can search its own past conversations.
• Skills it writes itself. After it works through a hard task, it saves the procedure as a reusable skill file so it doesn't have to figure it out again next time.
• Lives where you already are. One gateway process connects it to Telegram, Discord, Slack, WhatsApp, Signal, email, and the CLI. Same agent, same memory, whichever app you message it from.
• Runs wherever you put it. A laptop, a $5 VPS, a GPU box, or serverless sandboxes that hibernate when idle. Commands can run locally, in Docker, over SSH, or in a handful of hosted sandboxes.
• Scheduled work and subagents. It has a built-in cron scheduler for things like nightly reports, and it can spin up subagents to work in parallel.
• MCP support. Any MCP server you already use can be wired in as a tool.

What it's good at

You pick the model. This is the big one. Use a frontier model for hard tasks, something cheap for routine ones, and a model running on your own hardware when you don't want anything leaving the house. If a better model ships next month, you switch with one command and keep everything else.

Your stuff stays your stuff. Memory, skills, and config live on a machine you control. If you stop paying someone, nothing disappears, and the MIT license means nobody can take the code away either. It uses a variety of different memory providers. That's a bit too much for this post, but worth investigating if you decide to install.

It's at home on a server. Because it can run commands over SSH or in Docker, it fits naturally into a homelab or a dev box. Things like "check the backups every night and message me if something looks off" are exactly what the cron scheduler and messaging gateway were built for. Mine is running in a VM on a Proxmox host.

It gets better over time. I think the skills loop is the most interesting idea in the project. Most agents start from zero every time. Hermes records what worked, so the second time you ask for something specific, it already has a playbook.

It's hackable. It's open source, it speaks MCP, and its skills follow an open format. If it can't do something, you can add it. I'm always up for anything I can customize.

The new neighbors: Muse and Dots

Meta Muse launched September 8 in the US. It runs on Meta's Muse Spark model inside what Meta calls a Muse Secure VM, a dedicated cloud machine per user that holds the agent, your data, and the credentials for services you connect. You talk to it in the Muse app, on the web, or in WhatsApp, and it can book travel, fill out forms, negotiate, and check out with a one-time card through Stripe's Link. It's free for most uses, with paid plans above that.

OpenAI Dots recently showed up. A dot is an always-on agent inside ChatGPT, running on GPT-6 Astra, with its own cloud computer and browser and access to 4,000+ apps through OpenAI's plugins. This one is exciting because, well, Astra! When you're not talking to it, it does read-only "research" in your connected apps. I'm not sure if I love anything about that or not.

Side by side:

Why it might become less relevant

Most people don't want to run a server. A big part of Hermes's early appeal was that it was the only way to get an agent that remembered you, worked in the background, and lived in your messaging apps. Muse and Dots now do all three, and setup is "download the app" or "flip a switch." Non-technical folks will most definitely migrate to the button push options.

The big platforms already have the distribution. Muse is in WhatsApp. Dots are in ChatGPT, Slack, and Teams. Hermes has to be installed, configured, and connected to a bot token before it says hello.

The polish and safety layers are expensive to build. Meta's Sentinel approval agent, audit trail, and protected checkout, and OpenAI's monitoring and Agent 365 governance, are the kind of thing a company's IT and security people want to see before they approve anything. Hermes? That's on me...or you.

Its headline features aren't unique anymore. Persistent memory, background work, multi-app reach: those were awesome in February. Now? Nothing special. This AI junk moves quickly...

Why it might not

Every big-platform agent is married to one model. Muse runs Muse Spark. Dots run GPT-6 Astra. If a competitor ships something better, or the vendor changes pricing, you wait or you start over. Nous is betting that the durable asset is the harness, meaning your memory, skills, and workflows, not any particular model. I don't hate that.

Your agent's memory is a very personal database. It knows your projects, your habits, and eventually a lot about the people in your life. With Muse and Dots, that lives in someone else's cloud.

As I've recently learned, local models are getting good. If you've got a capable machine running a decent quantized model (see my GGUF post), Hermes is an obvious way to put it to work.

A dot's computer isn't your computer. OpenAI gives each dot a cloud machine, and you can optionally connect your laptop. But if your work lives on a homelab, a build server, or a box behind your firewall, an agent that SSHes in and runs your own scripts is a different tool entirely.

My guess is Hermes ends up as the Linux of personal agents: a smaller share of users overall, but the default among people who actually build things.

As far as I am concerned

If you want an agent that books your dinner reservations and you already pay for ChatGPT or live in WhatsApp, use Dots or Muse. I think that's going to be most people.

Hermes stays relevant for a specific crowd:

• Developers and homelab people who want an agent working on their machines
• Anyone running local models who wants them to do more than chat
• People who care where their agent's memory lives
• Anyone who refuses to be locked into one vendor's model

I'm in pretty much every one of those groups.

Hermes is the DIY hackable option which I will pick every single time. Anyone else?