>>109604161
Vibe coded version of Claude Code, modified to use less tokens and be provider agnostic.
>Antigravity
Dude Gemini-Cli was so so much better, Antigravity is a gigantic pile of shit.

>>109604161
>harness
I give mine access to an SSH tool uses screen to give it a "display" in one of my virtual machines. If it needs more than that it's too retarded to work on anything.

I use Opencode because the free deepseek flash is all I need

harness is the next browser war in ai era

>>109604161
>Grok BOTS!
???

>>109604161
Bondage harness
/thread
/locked

>>109604161
>>109604197
>>109604224
>HaRnEsS
It’s a software tool workplace anon get it right you dumbass.

>>109604161
opencode
i run it and it just werks with no fuss. i also have hermes but for that one i constantly need to mess with models and recently i started 429 errors on the nvidia API i fed into it. opencode just gives you a free api that i never managed to exhaust

>>109604161
>>109604188
>>Antigravity
>Dude Gemini-Cli was so so much better, Antigravity is a gigantic pile of shit.
its the only shit i have acess to. 3.1 pro is the slowest most retarded model yet. i just spam 3.7 high till it works. i have to make my prompts wit gpt tho js to be sure its not fucking up

File: late.png (93 KB, 847x584)

93 KB PNG

>>109604161
>https://github.com/mlhher/late-cli

what about abusing free limits of the webui ones

>>109604161
Why not use local LLM's and save on tokens all together?

>>109605608
((( poorfag )))
>>109605623
even with high ram ,laptop willl melt, no gpu

>>109605629
>Laptop will melt
What about a MBP or a liquid cooled gaming PC?

>>109605731
>poorfag
>mbp
nigga what about the avereage thinkpad an avereage /g/ user uses

>>109605623
because theyre garbage on anything affordable

>>109604161
Which of these and what AI (non-local) do people use to edit configs on Linux?

>>109604161
What's wrong with whatever come with Visual Studio or Code?
Works fine for me.

>>109606568
copilot?

>>109606653
I guess, dunno. I get it free at work so never checked.
It does have all the models you'd expect and the "harness" does it's job fine. I'm not big into agent mode, just use it and apply it's suggested changes one by one. Generating more than a few thousand lines of code gets iffy anyway, ai starts to forget things.

i just use openclaw

>>109604161
>harness
leather, of course

>>109604161
I found pi to be the best honestly. minimally intrusive + fast + still the most effective. Just hookup the MCPs that you need and let it run. Works perfect. I've had my runs do few hours at a time, before the other clients like opencode would stop after few tens of minutes at max. And the pi setup isnt just stalling, its actually working out in real productive work too. So its pretty good.

>>109604161
I use Pi, straight inside Emacs.

>>109604161
claude code in kitty in hyprland using caelestia with a bunch of custom scripts for managing sessions.
Each active project has a separate tab w/ notifications and also manages running their related processes and other stuff. I tried herdr at first (because I love the philosopher) and it was closer to what I wanted but I just pretty easily vibe coded what i actually wanted into my actual desktop environment and if I want to modify it at all it's effortless

Mistral Vibe works well enough.

>>109604161
Codex and OpenCode.
Played with Openclaw for like a month. Seemed like it was a neat tool, but annoying af because it was very much trying to use something while it's still being built.

>>109604224
Most of them do different things though.

>>109609487
Also, everything I'd get Openclaw to do, felt like I should just be vibecoding something that behaves more deterministically.

>>109605623
hermes uses local llm and you can design edge to cloud
>>109606457
you're a fucking idiot who doesn't even remotely understand this technology.

>>109604161
I use one I personally developed called my brain and studying. You might not have the hardware to run it though, fleshware requires years to tune and grow in ideal conditions, while also degrading if fed an information diet that includes ai generated slop.

>>109610064
>I use one I personally developed called my brain and studying
good luck with that, buddy

>>109610022
buddy, I ran glm4.7 32B on my 20gb vram gpu and the quality of the output was atrocious

>>109604222
i'm so sorry anon
now that they removed it, what do you plan on using now?

>>109610407
You will be denied entry to the kingdom of heaven once founded by man if you let your soul rot from ai exposure

>>109604161
>>109604224
>>109610022
>hermes
buy an ad vmfunc
>>109604188
fpbp fuck google so much fucking retards, still use antigravity because my job pays for it and won't pay for claude

Opencode. Considering something minimal, I don't like its bloated prompt with garbage like telling the model not to write evil code. Fucking faggot wasting tokens

>>109610512
i dont think that was in the bible

>>109604161
antigravity. flash is actually a pretty good model. its speed makes it easy to iterate and bugfix and you get a lot of use out of it before approaching limits

reminder that hermes is a project by a bunch of cryptoscammers and they advertise on this board

>>109604161
it feels good to be purposely ignorant of all this retardation, I've learned that AI subhumans use the word harness today, it's too much knowledge already, please delete your thread and use a trip so I can filter your future threads tyty.

Opencode is really nice. I remember not really getting along with hermes the last time around.
Claude code is like F tier.
Also, where's the copilot harness which was getting decent the last time around?

>>109604161
w-what happened to the SS tier?

>>109604161

The concept of the PTY harness explain many things to me. For example, if I am interacting as a user, my behavior is completely digitalized. That implies is possible to create a completely fabricated computer user.

I know sometimes during the operation of a classic determinist program is necessary to give a program a user account (for example Pulseaudio, X.org, Xlibre) but is like "flat".

Our interactions with the computer can be defined as a series of sockets and file descriptors.

>cursor
>the original that started all of this

So that was cursor, I see. The other service/program people talked about when cursor was all the fuzz was Repl.it I suppose that thing is also some kind of harness. Because those two where the most popular ai artifacts a couple of years ago and everyone was talking about them.

Cursor fell into obscurity because Anthropic presented a more eyecandy interface with that thing they supposely wrote with React Ink and supposely got "leaked". Repl.it collapsed after the incident when an unidentified agent or llm deleted a whole production database and people started to talk about MCP as the solution for that kind of disaster (please compare that real incident with the "an llm breached containment" hype/meme/panic)

In the meantime, IBM was pushing RAG.

The number of ESLs posting in this thread is telling.

>>109610476
Well, they have other free models but I don't know if they are as good as deepseek

>>109610831
the use of the word harness itself = esl

Ideally I'd like something like pi but compiled, with addons maybe being written in LUA.
A TUI in javascript is incredibly niggerlicious

>>109611291
write it faggot, i'm sure the 200mb ram saved will be worth it

>>109604161
>harness
stupid term

>>109604161
Hermes is a pile of shit
Pi is a chore
DeepSeek harness just works but it's barebones

>>109606547
???

>>109608179
the last woman using this in aug 2026

I’ve been using Cursor for the last six months and am very happy with it. When I see all the kvetching about how Claude ate all their tokens or OpenClaw rm -rf’ed their machine I can’t help but think I’m getting away with something. Shit just werkz. Grok Cursor gets a bit hungry sometimes but regular old Composer does about 80% of what I need and is cheap as hell on token usage.

>>109604161
Why do ALL OF THEM have telemetry?

>>109604161
I mostly use Hermes these days, although dsh shows promise with the plugin idea

>>109611558
You are the product. You certainly aren't covering costs with your token payments.

>>109612743
>You certainly aren't covering costs with your token payments.
you fell for the psyop. I ran the numbers for deepseek flash
there are two ways to run it locally without quantizing: two rtx 6000 (fast) or a macbook with 256 gb ram. either of this option is cheaper to run than the original "too cheap to care" deepseek flash inference at any provider.
a whole gpu rack that runs on HBM is a *lot* more efficient at inference, and you not only get batching but also dynamic batching where you delay and collage a number of requests so it's cheaper still.
tokes are borderline free, even from those huge models. the only reason why openai/antrophic aren't wildly profitable is their continuous costs of salaries + further training, not the cost of inference.

squeeb

>>109612743
>You are the product.
The person that benefits the most from training data is the user that submitted them.
It's so annoying to see this canned response trying to turn the idea of combining efforts into something negative. Why wouldn't I want the software I use to improve, especially when it's based around real world data of my real world use case. I'm the one that uses it, go ahead and tailor it for me.