AI DevOps + OpenCode - Alternative to _Claw bots

[marcusquinn]

I prototyped a Cloudron app for this here:

• https://github.com/marcusquinn/aidevops-cloudron-app

(made with aidevops)

you can of course still use it safely, locally — which I do and recommend — but if you want a cloudron instance for any reason, this should also work

look at my GitHub activity if you want to see what it's capable of:

• https://github.com/marcusquinn

[marcusquinn]

Massive amount of work has gone into token-optimisation with this, now.

Give it a try, and compare to whatever else you're using. I think you'll be pleasantly surprised

[LoudLemur]

This post is deleted!

[jdaviescoates]

Again, it's already packaged, I think.

I'm on the verge of reporting all your reports as spam @loudlemur

I get that you are coming from a good place and trying to be helpful, but imho you're just creating noise and these reports are not helpful given so many of them are obviously wrong.

[LoudLemur]

@jdaviescoates I shall hold off posting any more of these till there is some consensus on whether they might be made useful.

sorry about the noise!

[marcusquinn]

@LoudLemur don't litter external links, either. Not "helpful".

[umnz]

This looks cool. What are the chances of getting this officially added to Cloudron?

[marcusquinn]

@umnz i use and develop is daily. checkout my open:closed issues & prs ratio on the repo:

• https://github.com/marcusquinn/aidevops

(works with claude max subscriptions, too)

You can see my gh activity and ai-leverage throughput with it here:

• https://github.com/marcusquinn

There's a Cloudron app version here, but you're best off installing it locally first, and then getting it to walk you through setting up the cloudron app instance from this repo:

• https://github.com/marcusquinn/aidevops-cloudron-app

Maybe that can be added to the cloudron community apps library?

[umnz]

@marcusquinn really nice work Marcus. We're only allowed to use locally hosted models using OpenAI compatible endpoints. Does AIDevOps support local models? We have 10 or so nodes that can run anything from 24-70GB of RAM + context each on top (so most 30b models are compatible with big context). Any models, system prompts or settings you'd recommend for using this with locally hosted setups? I do like the sound of spooling up a bunch of workers and getting around the network limit by having them point to different nodes.

[marcusquinn]

@umnz OpenCode is the interface, so you can use any models, cloud or local.

aidevops is just a curated harness, to save everyone from reinventing the same, and it's designed for token-efficiency, speed, and optionality — with pulse & workers to get things done in the background, while your main interactive sessions can ask for anything

also, i've recently switched from using anthropic models to openai's gpt-5.5 (no thinking, ctrl-T until that setting is blank) and finding that very fast and reliable — i think with 5.5 they've finally surpassed opus level coding capabilities

for self-hosted models, i understand deepseek and kimi are the current SOTA for coding, let us know how you get on

[robi]

GPT 5.5 is roughly at par with Mythos.

[LoudLemur]

@robi said:

GPT 5.5 is roughly at par with Mythos.

That is interesting to know. Have they been using GPT 5.5 for pen testing?

People are raving about Hermes Agent. Would that fit into the AI DevOps framework?

[robi]

@LoudLemur Yes, they also ran other bench tests.

AFAIK No, Hermes is a different UI and harness.

[marcusquinn]

@LoudLemur Hermes is a bootstrap harness, bring all your own agents/skills, and no deep Git collaboration.

AiDevOps is an opinionated with 1000+ agent for things I use and recommend, and the only harness as far as I know that can work for multiple users and machines collaborating on the same Git repo (more difficult than you'd think if you try to do this without AIDevOps)

You can install all of these things, try them and compare.

Once you have installed an an account connected (GPT-5.5 is currently best for coding & devops), then you can ask it what it does, and to compare itself to anything else you find trending that gets your attention.

[LoudLemur]

@marcusquinn Thanks.

What do you think of this guy's take? He thinks the best agent is a markdown file and scripts that are routed through the markdown file, and these are best organized in ... folders...

[timconsidine]

@LoudLemur this is a better vid from him - long-winded but clearer :

[robi]

@timconsidine @loudlemur That is exactly what AID does. It works.

[timconsidine]

@robi
I don’t have a use case for it
I was just pointing those who might be interested to a resource

[marcusquinn]

@LoudLemur folders are a 1-dimension construct, so not always applicable to multi-use things

i prefer flat file structures where it can speed up searching and navigating for things

these are minor details, but i do pay attention to them

best way to see how i've structured and optimised things is to look at the codebase and try it

i'm at the point now where i use it to build more, so the evolution is now iterative based on that experience

i don't think you'll find a more token-efficient way to work at scale and get more done, but you're not locked into anything open-source, so try them all and compare

aidevops is good at answering questions about itself, and comparing itself to any other repo links you give it, it's a full-transparancy self-aware system

right now, the best model by far for coding is gpt-5.5 (no thinking needed, so Ctrl-T until that yellow text doesn't show in the opencode input) — pretty-much gets everything i ask of it done in one shot