Swarogan

AI engineering, founder

[Swarogan @swarogan]

For a while now I've been building products with Hermes agent, but I had no way to actually test them. The only plugin I found didn't work on Wayland, or on remote virtual machines.

So I wrote my own.

Now the agent gets a live view and full control - the local desktop or a Proxmox VM. It logs in, clicks, types, tests the app on its own, and I watch what it's doing live in a side panel. When I want to take over, I click the preview and drive it like a remote desktop - mouse, keyboard, everything. Let go and the bot carries on.

You can keep an eye on the local machine and multiple remote ones at once, and Hermes desktop switches the preview by itself to whatever the agent is working on at that moment.

It works the same on Wayland and X11 because it talks to xdg-desktop-portal instead of guessing the display server. That’s why the other plugins fell over.

I'ts open source. Link in the replies.

[Swarogan @swarogan]

@slimbaron Local models are great, I use them for translation, automation, transcription, and many other things.

[Swarogan @swarogan]

@brianmcgrath I used the CLI for a long time because the desktop wasn't convenient for me. But since Hermes Bots, it's completely changed my approach, and now I only use the desktop

[Swarogan @swarogan]

If you were building a new product today, how would you get your first 10 customers?

Communities
DMs
Paid ads

what would you try?

[Swarogan @swarogan]

Not noVNC. The plugin has its own async RFB (VNC) client written in Python that runs in the Hermes gateway, so there's no noVNC/websockify glue to set up.
I updated the Setup section on github, because it might have been unclear. I'm preparing a video tutorial, maybe I'll be able to upload it tomorrow

[Swarogan @swarogan]

I improved the communication, now it works with any novnc, not just proxmox, and runs smoother

[Swarogan @swarogan]

For a while now I've been building products with Hermes agent, but I had no way to actually test them. The only plugin I found didn't work on Wayland, or on remote virtual machines.

So I wrote my own.

Now the agent gets a live view and full control - the local desktop or a Proxmox VM. It logs in, clicks, types, tests the app on its own, and I watch what it's doing live in a side panel. When I want to take over, I click the preview and drive it like a remote desktop - mouse, keyboard, everything. Let go and the bot carries on.

You can keep an eye on the local machine and multiple remote ones at once, and Hermes desktop switches the preview by itself to whatever the agent is working on at that moment.

It works the same on Wayland and X11 because it talks to xdg-desktop-portal instead of guessing the display server. That’s why the other plugins fell over.

I'ts open source. Link in the replies.

[Swarogan @swarogan]

@mirbacx @CryptoCyberia If you're coding by hand, you're not delivering on time these days. I have at least three bots running: a coder, a verifier, and a supervisor.

[Swarogan @swarogan]

@AlexFinn @openclaw Same here. Too many bugs. I'm left with Hermes and Grok Bot

[Swarogan @swarogan]

@ericzakariasson @bot I want a working Linux version. Now I have to manually rebuild each update

[Swarogan @swarogan]

@mark_k OpenClaw 2.0 has a lot of bugs and doesn't work, I'm left with Hermes and Grok bot

[Swarogan @swarogan]

@aibyandres @bot @claudeai Treat the bot like an employee who calls the codex and you will have longer

[Swarogan @swarogan]

@RoliumGens Same. Qwen 3.8 27B dense daily, in Hermes.
Mail agent, local automations.

[Swarogan @swarogan]

Grok 4.7’s biggest advantage may not be the model itself.

It’s 25 years of SpaceX engineering data that competitors cannot scrape from the web:
design decisions, failed tests, manufacturing problems, autonomous systems and launch operations.

Initial training is complete. Supplemental training on SpaceX data is underway, and according to current announcements, Grok 4.7 should arrive within the next 2–3 weeks.

Having data from things that were actually designed, built, tested and sometimes failed may be more valuable than simply training on more internet text.

We’ll soon find out how much of that experience a model can actually learn. I’m really looking forward to testing it.

[Swarogan @swarogan]

@DogukanUrker I tested now Qwen3.8-27B-GSQ-RCO IQ3XXS and as you can see, it is weaker than Q4

[Swarogan @swarogan]

@ByanJati @xikhar No, I don't use Fable 5 much, tokens disappear before my eyes. With Claude I mainly use Opus 5, and sometimes Opus 4.8 because it doesn't have that awful censorship

[Swarogan @swarogan]

@OriSilver Done

[Swarogan @swarogan]

@DogukanUrker True. Gemma4 has a problem with calling tools and can get stuck in a loop

[Swarogan @swarogan]

@MichaelGannotti @NousResearch True. I started using Hermes Desktop seriously when bot mode was added. Previously, it was only available in the console. But I added a live view of the local or remote computer, allowing you to see what the bot is doing.

[Swarogan @swarogan]

Since the bots mode was added, I've been using different bots with different models. This allows the local model to handle sensitive data, and when it needs more power, another bot uses frontier models.
I created a plugin that allows bots to run on a remote computer, with a live view like Grok Bot, and it works even better