openclaw

Äm, i personaly do not want an AI agend on my pc. It is to dangerous for me.

Yes, very äm.
Well, you decide where it runs on, can be a server (though it should be protected well), or a seperate device.
Also tbh, it's only as dangerous as you allow it to be. And I'm not talking about promting, but just proper restrictions, and not just giving it access to everything you can think of. I have mine connected to various useful things, including Elten, actually, but I'd never ever do my main bank account (I could imagine if it had some kind of restricted, seperate one with a 100% enforced limit etc maybe), and not E-Mail either, as it includes some important stuff. And pw manager is a no go too of course. But for the rest, it also just depends which models you use. If yo use a local model, where is the problem with it accessing your messenger? Yeah, I wouldn't want Claude to read my WhatsApp chats. But if I'm looking for something or want it summarised, why shouldn't a local model do it?

Well it's a good thing hermes, I've noticed, has a refuse and flag design, where it will refuse to do distructave stuff if you say, plus it will put it in memory.

Hermes' "smart approval mode" is a joke.

I meen if you use it correctly you should be fine. But yeah I'm not a fan of the clarify tool timeing out for the approvel. But that's how it's designed I guess.Hermes' "smart approval mode" is a joke.
--

so question. This sounds cool and I would like to use Hermes to help manage my server. Would it be best to run on the server itself or on a different machine or how should I do this?

Hermes can handle both. If it'll be for ever editing text files and things on the server, it's better to run it on the server otherwise it's doing hacky things over SSH all the time to get things done. What you can actually do is set the terminal backend to SSH and specify your server, so any terminal command it runs is only on the server. I can see it trying to use its file tools on files though, so if you want it to do server things and mostly server things, may as well stick it on the server at that point.

what about the model itself, what should I use for that? My server does not have a graphics card so I can't imagine it would be able to handle much. My server is. root server vds 4 cores 8 gb ram 512 gb ssd

Well, I've connected Claude Pro sub, but I guess you can also get away with something open ai compatible like Nano GPT or anything they connect to.

Nothing decent in terms of local models is running on that server. I'd personally go for either Claude or Open AI, although other offerings like minimax and moonshot are actually pretty good. Minimax recently released minimax m3.1 flash which looks like it might be pretty decent. At least it's showing up on Ollama's cloud. Ollama's cloud might be what you want actually, you get access to loads of models and you get a certain amount of credits a month you can use and the model you use influences how far that gets you. So far it looks like minimax m3.1 flash is really cheep, also glm 5.3 flash is pretty good and not too expensive.

what about abacus chatllm?

I've tried messing with OpenClaw and alternatives but honestly I don't really see it being useful for me, it's a pain in the ass to set up and maintain and if you don't set it up correctly it's a massive security problem. I'm curious what others here are using it for though.

Mainly just research when I'm to lazy. It's running on a server, but since I'm connected to it through remsound it's one wake word, so when I'm lazy I let it search something up for me, send me the results on WhatsApp or something. Have it connected to a bunch of stuff like Elten, Telegram etc too just for the heck of it, but it's one of these things where you can argue you really don't need it. But it's fun to have and mess around with.

Prompt injection essencially was solved and I highly doubt that I am being targetted specifically. Call my naive or silly but I turned off smart mode or anything thereof. The only thing working as a gate is Hermes itself, which disallows doing sensetive things like sharing .env file anyway.
Also, here's a fun one: I asked Hermes to adjust my config. It said basically the following:
I couldn't edit config directly, so I made a script that does it.
If it really wanted to or really got jailbroken, it just could evade these safeguards anyway.

@34 I didn't mean prompt injection per say. What I meant is when you let any of these tools access your computer/server, you're essentially allowing another, very much so unpredictable, mind to access your machine and I just don't feel comfortable letting it have that access 24/7 without my supervision on a machine I care about. I do think the tech is very cool though and I've actually been meaning to set up my old laptop to be a machine just for running AI agents but I didn't get to actually doing that yet.

Ah yeah. I run it in a vm as well.

I have my openclaw doing the boring crap I don't want to do myself, like checking my timetable every night and checking for room changes and notifying me if anything changed, research for me as well, help me make university powerpoints and stuff like that accessible, that sort of thing. For me it's a very useful thing, and I can just talk to it on Discord and even send it voice messages when I'm away from the computer.

and which mode runs your openclaw / hermes? I prefer hermes since I am not a terminal person and the desktop app suits me well, I currently have it running through TG.