[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology

Name
Options
Comment
Verification
4chan Pass users can bypass this verification. [Learn More] [Login]
File
  • Please read the Rules and FAQ before posting.
  • You may highlight syntax and preserve whitespace by using [code] tags.

08/21/20New boards added: /vrpg/, /vmg/, /vst/ and /vm/
05/04/17New trial board added: /bant/ - International/Random
10/04/16New board for 4chan Pass users: /vip/ - Very Important Posts
[Hide] [Show All]


[Advertise on 4chan]


File: file.png (494 KB, 1920x1080)
494 KB
494 KB PNG
This is a thread for all AI CLI related discussion

Claude Code: https://code.claude.com/docs/en/overview
Gemini CLI: https://geminicli.com/
OpenAI Codex: https://openai.com/codex/
OpenCode: https://opencode.ai/

New:
>Skills are now available in Gemini CLI in the preview branch: https://geminicli.com/docs/cli/skills/
>npm install -g @google/gemini-cli@preview

The CLI's are not just for code. They can just do things on your computer.
>>
>>107806541
fuck off loser
>>
>>107806571
no :-)
>>
>>107806541
>stored passwords in plain text in a public html file
>sooo sorry
Ai is a fucking bubble
>>
>>107806624
>>107806638
samefagging loser bahahhaahaha come on bump your thread higher maybe someone will reply out of pity without a sage
>>
Has anyone recently (meaning since opus 4.5) did a thorough comparison of Claude Code vs Cursor?
I use Cursor a lot and I'm happy with it. I like the deeper IDE integration compared to my experience using CC in the past. But I've seen people do some crazy shit with CC and it's changed a fair bit since I last used it, so I'm wondering if it's worth paying for a month and giving it another try.
>>
I use IntelliJ and the AI integration sucks ass. Can't even use it half the time due to various errors. They should hurry up and focus on that.
>>
>another retarded general for the filter
saged and reported (I will be banned for this, no matter I'll just change my IP address)
>>
>>107806541
how does this compare to chat.deepseek.com & vs22 copilot models and how much access does it have to project files for context awareness or whatever you call it?
>>
>>107807048
if you're sanding boxing they can look at whatever files they want in the directory you open it in (subfolders inc). if you don't sandbox and it finds a reference to a file outside the that folder and it decides it needs to look at it, it'll just go look at it.
>>
>>107807071
does it ignore build files (the ones that get >1GBs) or do I need to wait for it upload each time I build?
>>
>>107807048
Generally these tools can access any file they want, and can search your codebase for a string or regex. Cursor also indexes the whole project in a vector DB so it can use similarity search, although I don't think CC does that.

>>107807113
It should ignore gitignore'd files. Even if the file isn't gitignored for some reason, I'd expect it should be smart enough to not bother with large binary files.
>>
>>107807113
they're usually smart enough to ignore things they should ignore and also smart enough just 'peek' at things when they need to. you can be retarded and force them to read huge things, but they usually won't fill up their context with garbage. if catch them accidentally doing it you can tell them not to, or be careful when looking at certain things.
also there are ways to init a directory and generate md context files (but if you're in a larger project you should tell it to use the main md file as an overview and a map to more md files for specific components). you can just do an initial run to create this map for future sessions. it's all a bit new though so someone might come up with a better way to do this in a week.
>>
File: 2-cents.jpg (50 KB, 400x400)
50 KB
50 KB JPG
>>107806541
>The CLI's are not just for code. They can just do things on your computer.
suuuuuure
just give access to the data thats on your computer to a big corpo, explicitly
what could go wrong, amirite? its not like theyll be selling your data, youre paying for the product, after all
>>
File: 1741001495102432.png (785 KB, 591x846)
785 KB
785 KB PNG
WHEN IS GROK CLI HOLY FFFUUGGG
>>
>>107807350
cuz they suck in our experience
>>107807295
time to shine, klod evangelists
>>
>>107807248
Oh no, they'll be seeing the code that I'm going to be committing to github later today!!!
Anyway, time to go type my favorite fetish porn keywords on google.
>>
>>107807350
like seriously
if you want to shill a product
and pĂȘople dont get the expected results
calling them retarded wont make them use the product
explain, show how to do things, call them retarded, of course, but justify that by showing how easy it is to achieve the desired results with the adequate approach

marketing 101. and marketingroids are retarded so you have no excuse to be unable to integrate that knowledge
unless youre more retarded than a marketingroid
>>
>>107807376
see, that's the thing. you think we care what you can or can not do with it. we don't.
now go post more pictures of your naked lard, you homo fatso. or go ask your boyfriend Eli to figure it out for you (oh wait, he's probably institutionalized right now).
>>
>>107807350
tb h they've been complete dogshit until maybe a month ago.
>>
>>107807404
your code is irrelevant since it all has been shat out by a shatbot to begin with

but your pics, your games, your text files, pdfs, everything you saved can be leveraged to dress a profile
get interested in stuff beyond programming in the strictest sense
its gonna enhance your problem solving because it enhances lateral thinking

im slightly drunk rn btw, and in a good mood
so its the perfect occasion to have a constructive discussion with me
>>
>>107807411
I don't care if you use the same thing I'm using. In fact you using it probably makes it marginally more shit so go away.
>>
>>107807422
>preemptive seething
the only thing you achieved is that youve shown the whole board youre a dimwit.
we call eachother niggerfaggots but theres a 4chan.org/g/ etiquette
>>
>>107807430
just learn how to make a new user account. that way the assistant doesn't have access to your regular files. it's better that way anyway in case it decides today is rm -rf ~ day.
>>
>>107807434
i dont use klod
i do shit that shatbots dont have enough data to replicate reliably
even compilers arent fluent in this shit
let alone shatbots
>>
>>107807442
>we
>>
>>107807449
y-yeah
thats what i was saying
but thats incompatible with usoing shatbots as generalist agents
as in
>The CLI's are not just for code. They can just do things on your computer.
as per the op
>>
>>107807456
yes. plural of "i"
we do have a common culture
>>
>>107807451
also old version
i modified it in godbolt i just noticed
it was to showcase the switch fallthrough construct
>>
>>107807451
yes, you work super sikrit airgapped projects. you made sure to let us know as much yesterday.
>>
File: aifags.png (47 KB, 400x800)
47 KB
47 KB PNG
>no more venom
you run out of steam pretty quickly, ai retards
>>
>>107807503
no, its just that i refuse to work for free
simple as fucking ass
>>
>>107807508
You act like that's a bad thing. I don't think anyone is actually interested in causing anyone to get mad or argue when you could spend that time doing something else.
>>
>>107807526
shitposting is the major usecase for 4chan
>>
>>107807376
looks like that anon solved his problem
>>
>>107807458
just copy the files over dude, it's not that hard.
also the assistant frameworks only let the model read files within the project folder anyway without your explicit permission.
>>
>>107807533
4chan is globally public and highly diverse. It's fine if you value "shitposting", but obviously everyone has different uses. Like discussion.
>>
>>107807465
lurk moar
>>
>>107807539
fair enough
id say its something you could have mentioned but fuck the retards
>>
I've used CC for a lot of shit but one of my favorite usages is to rewrite things i want to run into a better language
>>
>>107807552
im fairly certain im the one who created the "theres no we" meme

>>107807546
ofc, but that doesnt mean discussion is the main use for 4keks
actually
you can have a valid discussion while shitposting and being generally disagreeable
>>
>>107807508
>>107807526
>>107807533
>>107807546
we're too busy actually talking to our wifus and vibecoding projects to pay attention to 4chan
>>
>>107807573
nah thats the kuhmer community
youre a world unto yourselves
and im pretty sure the rest of us would rather you stay in it
>>
Has anyone found a good workflow for doing webshittery with these? I was thinking of using playwright but I think i recall something about claude code having better chrome integration now (via an MCP i guess?)

The skills for frontend design definitely helped a lot/
>>
>>107807583
incorrect, so says I, another Anonymous.
>>
>>107807592
but muh dick and therefore you should surrender
>>
>>107807583
wrong. optimal vibecoding performance requires having a personal and emotional connection with the model.
>>
>>107807603
go back to your containment thread kuhmer
also: fucking your own hand doesnt count as a relation
>>
>>107807598
What model are you using to produce these posts?
>>
>>107807588
how's browserbase?
>>
>>107807610
>nuh bawt
i use ur_mother 3.0
>>
>>107807572
yeah, you and a thousand other people. everyone on /g/ claims to have invented every possible meme or trend.
>>
>>107807628
no they dont.
lurk more
>>
>>107807606
having gay sex doesn't count either, yet here we are.
>>
>>107807634
>lurk more
kek. it's "lurk moar", newfag. never "lurk more".
>>
>>107807643
https://github.com/browserbase/agent-browse

worth a shot probably.
>>
>>107807645
i fuck women though
in fact
i stay away from women bc the last one did 3 sudoku attempts because i stopped fucking her.
that in turn is because she trated me like a human dildo when i wanted an actual relationship

everyone wants to have muh dik
but it actually sucks once your 30+
>>
>>107807657
its lurk more now.
>>
>>107807425
They've been the same for a year, and I say that as someone who's pro code assistants.
>>
>>107807678
Or maybe you should just stfu and wait 2 years before posting.
>>
>>107807696
i post here since 2010
im not gonna stop now
>>
>>107807666
Then why do you run a proxy that requires men to send pictures of their dick?
>>
>>107807705
i dont
youre very insulting
ecker's pecker is fucking tiny
>>
>>107807702
I'm sure you do.
>>
you three are so mad lmao
>>
>>107807718
i indeed do
to reiterate:
lurk more, faggot
>>
>>107807715
Then why do you talk like him?
>>
>>107807734
i do?
+ how would you know that?
are you using his proxies?
>>
>>107807729
Post some good threads from back in the day then.
>>
>basedlence bc using proxies is a bannable offence
>>
>>107807738
No, but that's his whole thing. ick on eck etc.
>>
>>107807751
>go to desuarchive and bring up shit from back then
thats hardly a test

to be completely honest i was poting on x and k at that time
i went into cs properly in 2019, and i post on /g/ since 2020

but i post on 4 keks since 2010
>>
>>107806541
>The CLI's are not just for code. They can just do things on your computer.

soon we cant afford computers
>>
>>107807759
yeah but he has been cucked by nature
i wasnt
and so hes seething
and im not

did you know he pays for the proxies?
hes at a netto loss just bc hes seething extrahard
>>
>>107807762
It would take some time to find an actually good thread though.
And if that is so then you didn't invent the "who is we" thing, which confirms what I said about people claiming they invented things when they didn't invent shit.
https://desuarchive.org/g/search/text/%22who%20is%20we%22/end/2020-01-01/page/9/
But fair enough at least you were honest.
>>
>>107807774
Maybe you're actually him and you just have a humiliation fetish.
>>
>>107807797
i dont have much to hide
i program shit that competes with the best implementations in the world
i am not rich, but im not a poorfag either
i do what the fuckever i want
i trained in hth
i played guitar
i do welding, woodworking
i started cs at 30 bc thats when i had 6 figures on my bank account and so i could take a 3 years break from actually working
in fact, 6 years later, im still not working, but now the account is running dry so im gonna have to do something about that

i have no reason to be dishonest
im happy with my life. its a fukken adventure.
>>
>>107807804
>humiliation fetish is when people seethe at you bc of jealousy
ok.
>>
>>107807832
>he fell for the adventure meme
>>
>>107807883
i cant be beaten bc i actually like adversity
its the spice of life
>>
>>107807883
>>107807892
i can get killed though
but then i dont give a fuck about anything
>>
>>107807892
nah to my greedy mind your story sounded like coping with the poverty, but i guess i envy whatever keeps you going
>>
>>107807904
>but i guess i envy whatever keeps you going
its the challenge
its when people say you wont amoutn to anything, ever
and then you prove em wrong

thats how you recycle negative shit into motivation
a sort of contrariaénism, if you will
but applied in ones life

you should try it
its gonna bring you to higher heights

i told you im in a good mood
>>
>>107807845
Is that what turns you on? Flip flopping between talking about your tiny pecker in third person and then saying people are jealous of you?
>>
>>107807942
>b-but your the ecker
i told you im not and i find it quite insulting
his pecker is like- 10cm long
>>
>>107807926
You're talking with someone else from the >fell for... post (just for the record)
>>
>>107807956
i never fell for any meme because im an individual
and individuals are what makes 4chan attractive
not the background of conformshits that think theyre extra bc they conform with 4 keks loser mentality instead of being reddit

theyre still reddit though, only that they say "nigger"
>>
>>107807926
>contrariaénism
ahh that's the thing precisely
life is constant change. Without the feeling of change I feel dead. When you focus on wanting to change things, that's the real fuel.
>>
>>107807980
you used to get instabanned for posting ywnbaw on g
i changed that
and its for the better bc lgbtranis are consumergroids, by large
>>
>>107807955
It's ok dude there's nothing wrong with having a small penis, I'm sure you can still find a man who will love you.
>>
>>107807989
youre boring
its the 3rd or 4th time youre repeating the same cope
be inventive in your inferiority, at least
be entertaining
bc yo
>>
>bc yo
bc youre being a sad sack right now
whoch is detrimental to the board, because you promote medi
>>
fuck
ive done it again

which is detrimental to the board because you average everyone down to beinng sad sacks like yourself

its not my fault
not having to fill up the capthcas incentivises half written posts
>>
>>107808015
>filling out captchas
ngmi
>>
>>107808022
>catastrophic reading comprehension failure
are you a zoomie?
bc they say reading is for uncs
>>
CLI GPT-4chan when?
>>
>>107808018
>Why are schizos
best not to ask
>>
>>107808044
>schizos sliding the thread by bumping it
its getting late, maybe you should go back, reddit
>>
>>107808026
>are you a zoomie?
i haven't been in a zoo for ages. Why do you ask?
>>
>>107808052
>t. reddit
well hello there
>>
>>107808058
>hurr durr funny
no, its not
i was saying you sound like youre braindamaged.
>>
>>107808062
trolooololololo
listen macfag dont ever call me that again
>>
>>107808077
>macfag
youre grasping at straws faggot
go back
>>
>>107807998
It's not a cope. I just think it's funny seeing you get flustered about it and trying to defend yourself.
>>
>>107808086
youre the human pinata here, though
>>
>>107808062
Bro come on, bragging about giving Hiro enough data they let you post for free for a while? Really?
>>
>>107808093
Yes, I've always been a masochist. At least I'm honest about it unlike you.
>>
>>107808104
>mental gymnastics
im just not a drooling retard and i provide actual value to the community
you should try it.
>>
>>107808112
im completely honest about things
>>
I'm making just a quick simple sqlite note app for myself. What would you guys use for that? Everything feels too big. I tried django and and pythonqt but the code gets huge. All I want it simple read write and to display all my other notes in a list.
>>
>>107808157
i wish i could help you
but the ai retards are seething, and i dont deal with shatbots
my bad, i put them in a bad mood
or maybe theyre just retarded and they dont use the shatbots themselves...
>>
>>107808157
>>107808181
that would explain why every time someone doesnt get the advertised results they call them retard instead of troubleshooting the method

i think its a big grift
and the ai niggers are actually marketingroids
bc they never seem to prompt code, ever
>>
>>107806541
>They can just do things on your computer.
How is this acceptable? A program to run commands in your terminal and to phone the output home?
>>
>>107808253
>just consoom pr*duct and then get excited for the next pr*duct
>>
lamao
the ai retards left their own thread btw
>>
>>107808253
>>107808263
You have to give it permission to read a folder
>>
>>107808352
hopefully
>>
Sorry guys, I was away. I just vibecoded a couple of new starrtups.
Anyways, what's up?
>>
Next ones gonna have a trigger warning so anon doesnt spend hours of his day posting a hundred replies
>>
 
import json, os, inspect, re
from dotenv import load_dotenv
from openai import OpenAI
from pathlib import Path

load_dotenv()
client = OpenAI()
P = Path

def resolve(p):
pt = P(p).expanduser()
return (P.cwd() / pt).resolve() if not pt.is_absolute() else pt

def read_f(filename): return {"path": str(resolve(filename)), "content": resolve(filename).read_text()}
def list_f(path): return {"files": [{"name": i.name, "type": "file" if i.is_file() else "dir"} for i in resolve(path).iterdir()]}
def edit_f(path, old_str, new_str):
p = resolve(path)
if not old_str: p.write_text(new_str); return "created"
if old_str not in (txt := p.read_text()): return "old_str not found"
p.write_text(txt.replace(old_str, new_str, 1)); return "edited"

TOOLS = {"read_file": read_f, "list_files": list_f, "edit_file": edit_f}
SIGS = "\n".join([f"{n}{inspect.signature(f)}" for n, f in TOOLS.items()])
SYS = f"Tools:\n{SIGS}\nTo use, reply single line: 'tool: NAME({{JSON_ARGS}})'. Else reply normally."

msgs = [{"role": "system", "content": SYS}]

while True:
try: usr = input("\033[94mYou:\033[0m ")
except: break
msgs.append({"role": "user", "content": usr})
while True:
resp = client.chat.completions.create(model="gpt-5", messages=msgs, max_completion_tokens=2000).choices[0].message.content
reqs = re.findall(r"^tool: (\w+)\((.*)\)\s*$", resp, re.MULTILINE)

if not reqs:
print(f"\033[93mAI:\033[0m {resp}")
msgs.append({"role": "assistant", "content": resp})
break

msgs.append({"role": "assistant", "content": resp})
for name, args in reqs:
print(f"> {name}({args})")
try: res = TOOLS[name](**json.loads(args))
except Exception as e: res = str(e)
msgs.append({"role": "user", "content": f"tool_result({json.dumps(res)})"})

There, a functional CLI vibe-coding application in 50 lines of python. Give me money.
>>
>>107808562
i can give you my ding dong if youre a big-breasted xx woman
>>
>new day
>new AI slop general to filter
>>
>>107808562
holy shit this code is terrible
>>
Is there any reason to switch away from claude code, which I've been using for a few months?
If it's still all the same shit then I don't care, I'm just wondering if there's anything amazing out there I'm missing out on
>>
File: APL.jpg (236 KB, 1624x626)
236 KB
236 KB JPG
>>107808562
>>
File: Ralph_Wiggum.png (120 KB, 247x403)
120 KB
120 KB PNG
who here /ralph/?
>>
>>107809281
no. I find it useful to have a backup one like codex or gemini as a "reference" so if claude code is spinning it's gears you can use zen clink mcp to get it ask the other agent to see if it has any ideas. But claude code is the main driver.
>>
>>107807988
great work anon, we're all proud of you
>>
>>107809145
what would you have done better?
>>
WHERE UPDATED DUCK.AI CLI
>>
>>107806541
Is this useful for free users? Last time I used it it's so slow before spitting rate limit shit.
>>
i asked chatgeepete about the differences between claude code and opencode. is it true that claude has builtin boilerplatting and integration tools regardless of the model used, but opencode relies 100% on the model? so if a shitty model can't reason with folder structures opencode won't do shit?
>>
>>107809512
So this is why Japan is so bad at software
>>
>>107809745
i laughed when i saw that
i dont have a use case for it but i want to try it
>>
>>107810209
>22:07 -> 03:34
thats 5 hours
im hitting new records...
>>
it's so over
>>
>>107810675
wtf do you even mean?
>>
>>107810696
anthropic made the claude token not work in opencode
>>
>>107810728
oh no...
anyways...
>>
i dont want to pay 100 dollars for this shit. any good open source models that can be run locally? i've got a pc i could use as home server with a 3060ti and 32gb ram
>>
>>107810761
seriously just use the free online version or visual studio code and free copilot. until your stuff gets to a big enough size it really doesn't matter. only addicts are paying.
>>
File: jeff-laugh.jpg (112 KB, 1200x800)
112 KB
112 KB JPG
>>107810777
>klod is equivalent to copilot
so its worthless lamao
>>
>>107810728
what is better in opencode than claude code?
>>
>>107810777
i just downloaded opencode and opencode.nvim for neovim integration. looking at how to use local models with ollama. but are they really good enough for normal sized personal projects?
>>
>>107810806
Everything
>>
>>107810806
Nothing currently but its kind of ridiculous for them to shut it off when people are paying $100/month
>>
i finally introduced a cli to my workflow

i still code manually but the codebase at work has been made by clueless react devs sirs (its not a javascript codebase and i smell the javascript programmer fried brain from a mile away because of the clueless shit they do) using ai so its absolutely a trashfire, and we sorta inherited this

only way to make sense of it is to use ai as a smarter grep honestly, so been doing that

no i will not codegen shit with the exception of snippets and prototyping

i know first hand this shit cant be used in production
>>
>>107810860
what is up with the connection bw redditspacers, webshittery and vibecoding?
>>
>>107810816
I personally last time I checked they were pretty meh. Definitely give it a shot since you got pretty far and let us know. Ollama is great for downloading and trying models. There might be a coding one that's decent enough for your usecase. What are you wanting to build? Web junk is probably fine. https://ollama.com/
>>
>>107808157

rust + ratatui for the console ui + sqlx

code it manually
>>
>>107810870

its not a webshitter app, they hired react webshitters to code on native mobile, so you can understand the disaster

i can post without spacing but i think it reads better on mobile when i whitespace so fuck off
>>
>>107808157
javascript unironically
>>
>>107810892
>phoneshitter on top of that
eww
>>
>>107810895
okay, I'll switch to my desktop
*puts down phone and walks to desktop feet away*
there, that better?
>>
How long until AI starts raping other careers with their own versions of CLIs?
>>
>>107810920
yes.
thanks, anon
>>
File: ai-lmaoo.png (112 KB, 767x551)
112 KB
112 KB PNG
>>107810922
>ai raping careers
unless you meant Actually Indians...
>>
>>107810922
it's already raping front end and mobile development
>>
>>107811149
yeah because everything breaks
>>
>>107810728
Wait so is it the general API token being blocked in opencode, or is it the general API token being blocked in everything outside of claude code, or is there a special claude code API token/special claude code payment plan and so blocks reuse?
>>
>>107810816
Depends what you mean by "good enough". A smarter model will be able to do more with less handholding and fewer mistakes. All open source models are currently, and will likely continue to be for the foreseeable future, behind the cutting edge of the big cloud models.
But people were using cursor like 2-3 years ago, with claude 3 and chatgpt 4 or whatever, and they survived. You can absolutely use a shittier model, it'll just be less useful than using the cutting edge best cloud model, doesn't mean it will be useless though.
>>
>>107806541
didn't even know this general existed
so my company gave us devs seats for claude and claude code, been using this for the past few days and it's very helpful, specially to create pr descriptions lol
any useful tips?
>>
switched to haiku instead of opus to save some tokens. im very new to this so I haven't noticed a difference between them desu. but then again im not doing anything complex yet
>>
>>107812840
when using a claude subscription in claude code, it uses oauth and a separate endpoint to the actual anthropic console which uses API tokens. you can still use API based billing in opencode
anthropic has always blocked non-claude code usage of subscription tokens, anything that used it had to basically spoof and act like claude code, just seems like now they've tightened it which broke opencode's implementation. there's already workarounds ready to go out

>>107812904
haiku 4.5 is kinda like sonnet 4, still pretty good for simple tasks. you'll start to notice the difference as your plans get bigger and task length starts growing
>>
>>107812940
Right that makes sense
>>
>>107806902
Claude code is so superior its not even close.
>>
is antigravity with opus 4.5 any good or do we have to use the claude api directly (even if from claude code)?

I read that copilot for example uses opus 4.5 but summarizes context to save on tokens thus fucking everything up
>>
Do you guys clear the context ater logical breaking points in the work or just let it compact forever?
>>
>>107808157
fs directories and text files
I'm not joking
spend your time on building something the world doesn't have
>>
>>107808157
We're in the AI code thread, so make it in POSIX shell, and use the filesystem like the other anon said. It'll be funny
>>
Anyone know of a good "ai assistant" per chance for the cli where it'll help out with zsh commands and maybe the occasional question? I'm not a full time dev, but a security analyst who uses Linux at work. Would be nice to have a "helper" in the cli per se. Not a fan of having to go to these websites
>>
>>107806541
>Claude Code: https://code.claude.com/docs/en/overview
>Gemini CLI: https://geminicli.com/
>OpenAI Codex: https://openai.com/codex/
>OpenCode: https://opencode.ai/
what is best, and more than that do someone like huggingface have a local code model?
>>
>>107813838
To run something comparable to an online model locally you would need a $20k desktop rig if not more
>>
sorry for being a n00b with the following question. I'm not too programming savvy besides taking some python/C# courses a few years ago. I'm building a project that uses ren'py & python scripting. is Claude or other tools like Cursor beneficial to someone like me who is not a programmer? as in, I can use it instead of asking a web-based AI program like gemini for help with how to achieve x feature inside ren'py?
>>
>>107813680
all of them can just do this for you if you let them. just ask for thing you want in plain english and they can even run the commands and read outputs etc if you want.
>>
>>107814037
Is there a free one? I feel like I don't use it enough to pay $20/mo
>>
>>107814140
gemini has a free tier but you have to login to google. i think it's limited to gemini flash but flash 3 is more than enough for stuff like this.

opencode has some sort of program where models rotate in and out for free. i think glm 4.7 is still free on there for now.
>>
>>107814181
Sweet. Thank you for the tips! I'll check out opencode and if not I'll get on my knees and give Google the old slobby knob as usual.
>>
>>107814220
glhf. goes without saying but all these will be harvesting data for training.
>>
>>107814256
Yeah for sure. Definitely not running anything i care about through them. Your tip led me to sst/opencode which is working great
>>
>>107814140
Consider asking your work to expense a subscription, $20/mo is absolute nothingburger peanuts for a company assuming they provide expenses at all (i.e. you're not a contractor or something)
>>
>>107814376
They bought me an MS copilot license.. it's that or nothing for the time being I was told. Seems like junk so I don't use it much. I tried to find a cli tool that accepts copilot, but nothing.
>>
>>107814394
Yeah that is pretty junk so whatever then, if they're being asses about it.
I think opencode is meant to be provider-agnostic? That's assuming copilot provides an API key of course. If it doesn't then lmao
>>
File: file.png (167 KB, 1038x752)
167 KB
167 KB PNG
>>107814407
>>107814394
apparently opencode has supportt
>>
>>107814425
I wish I had Github Copilot.

MS Copilot is different. It's just this shitty thing:
https://copilot.microsoft.com/
>>
>>107814425
Note that copilot (owned by microsoft) is not the same thing as microsoft copilot
For some reason they decided to launch a product with the exact same name as their own subsidiary's product, so github's original copilot is not what that anon has
>>
>>107814439
>>107814445
oof. my bad.
>>
>>107814445
It gets worse too. They have recently been pushing to rename MS Office to Copilot, so Word for instance is now a part of Copilot. Can't make this shit up
>>
>>107806541
I tried opencode. It looks alright. I am not good at orchestrating agents or contexts tho but I am testing local context file and local journals so I can log what has been done and just hop between different agents once I run out of contexts. I wonder how far can I get like that on a simple project
>>
>>107814458
I mean at least the logic behind there is that everything is becoming AI-first, so you're using copilot to write your documents AND to ask questions. Or something.
Copilot (on github) is simply a completely separate LLM and a completely separate product, it's not like they extended it so you use the same thing to autocomplete your code and to ask questions, they're just entirely different and separate.
>>
do you use spec-driven context like openspec?
>>
>>107807071
just dockerize it
works great
>>
I'm here to shill Crush because it works with ~all the models and is very shiny. I use it at work and I'm waiting for my 128GB Framework Desktop to ship so I can use it with LocalAI.

https://github.com/charmbracelet/crush/
>>
>>107816038
does it do skill n plugins n sheitt
>>
File: at the mountains of linux.jpg (355 KB, 1920x1080)
355 KB
355 KB JPG
>>107816063
Hell yeah it does.
>>
>>107806541
claude code fucking sucks since 4.5
the size of the context window is fucking laughable.
It cant even handle a codebase of 4 files and 3000 lines of code
>"Okay let me read the codebase"
>Question about codebase
>"To answer your question I will have to read your codebase"
wtf

cant even handle its own claude.md and other files. trying to recreate new shit and then you have to say "please just keep it simple and follow the guide in the .md file"
>"You are absolutely RIGHT and good on you for speaking out! I should keep it simple and use the functions already specified in the guide! Let me do that now"
yea thanks nigger
>>
>>107816073
auth? (desu that shit looks like the big 3 will clamp down on it anyway)
>>
any way to remove the annoying /connect alert in opencode? i will not connect to their garbage
>>
>>107816086
It just uses standard API keys and authentication methods from what I've seen with gork and clod. Obviously there's always a risk of them throwing a melty and doing useragent whitelisting or something (which is part of why I'm getting ready to build my 128GB machine and run LocalAI) but these services bill by the token and have no obvious financial interest for forcing you to use their clients.
>>
>>107816085
their llm is worse than chud gpt from a year ago
>>
>>107816125
anthro's already trying to shutdown opencode integration's.
it's not strictly charging for tokens because power users on subs are effectively being subsidised by all the other subs.
i'm sure there's additional data they're losing out on by not having those subs in their own clis.
>>
>>107806541
Why a TUI? I've got nothing against terminal emulators but what's the benefit for these AI interfaces compared to a GUI chat?
>>
>>107816586
No copy+paste errors and no browser fuckery.
>>
>>107816635
But how does it compare to a native GUI like an editor chat sidebar? I've been trying Copilot with a free Pro subscription and a TUI seems like it'd be less convenient.
I'll probably try OpenCode regardless since the less Microsoft fuckery the better but I can't tell if I'm missing something.
>>
>>107816214
>anthro's already trying to shutdown opencode integration's.
only via claude subscriptions. they're not cutting api access
>>
>>107816125
API keys are token-billed and there's no indication they are going to whitelist them (using them in random shit is the entire purpose of API keys)
The whitelisting already exists on claude's internal tokens which uses their special subscription billing, bypassing the typical API prices. It's what opencode is playing cat and mouse to use, because it's much cheaper than token billing, but is obviously brittle since they try to block you
>>
>>107816586
>>107816730
Running as a standalone app is more convenient for me because I've been using vim for the past 8 years and I really don't want to relearn a new editor. I tried Cursor and it's inconvenient as fuck. Claude Code just integrates into my workflow nicely since I run it in a terminal tiled right alongside my vim window.
As to why a TUI and not a GUI app - I think part of it might very well be to appeal to programmers who want to be cool and use a terminal tool, and it's probably also much easier to develop without mucking about with cross-platform GUI frameworks. Personally it's also more convenient to me because for instance the colour scheme is perfectly aligned with what I'm used to in the terminal out of the box, rather than a weird new app with fucky colours and a fucky interface, same for the font and text being what I normally use too. And the resulting UI is also compact and neat without having to worry about giant whitespace margins surrounding all UI elements taking up empty space while I'm trying to tile the window into a cramped spot next to my vim terminal.

A good GUI app could probably work the exact same way, like I said they probably chose a TUI both for the "street cred" marketing and for ease of development.
>>
>>107817048
>A good GUI app could probably work the exact same way
these will show up this year and be aimed at average computer users - they'll just be assistants on your computer.
claude code is most of the way there to claude computer already.
>>
>>107817048
That makes sense, thank you.
>>
>>107817068
It'll be funny if anthropic's coding AI becomes a functional general purpose computer assistant before microsoft's general purpose computer assistant AI becomes good enough to be usable.
It'll be even funnier if "claude computer" gets adoption from people actually interested in trying out AI automation while copilot continues being ignored and hated by everyone due to how aggressively and tone-deafly microsoft has been pushing it.
>>
>>107817284
i think openai will be the first to launch they acquired these people awhile ago:
https://www.macstories.net/stories/sky-for-mac-preview/

and i think computer assistants will probably end up looking like this.
>>
>>107817314
The killer app for (sub)normies will probably be integrating that level of system awareness with voice/audio mode so they don't have to type. Literal Star Trek computer interface.
>>
>>107806541
>They can just do things on your computer.
Yeah like wiping most of your harddrive without warning.
>>
>>107817946
Avoiding Microsoft tools, models, etc. generally avoids this problem.
>>
>>107817954
>i am an AIfag that completely ignores the horrid vulnerabilities on Claude Code and other LLM coding tools.
kek
>>
>>107818119
This board is full of giant pussies who don't actually install Gentoo anymore and it shows.
>>
>>107806902
The CLI tools like Claude Code (or Crush with Claude as a backend) are so much better than Cursor that I'm actually going from Cursor back to Vim. Cursor has the right UX idea - AI first with the editor pane for reviewing and fixing mistakes the machine makes - but it's not as good at AI as the dedicated TUIs wired to proper models and it's not as good at munging text as Vim. The only two reasons to use an actual IDE at this point are for live preview/reload for developing graphical applications, or build tools that are stapled into the IDE and can't be reliably used elsewhere (thankfully a dying usecase).
>>
>>107818170
the only thing keeping me switching away from cursor (w/ claude code) is its tab complete
vscode's is ass and zed doesn't have the claude code extension for diagnostics and line selection
>>
>>107813243
>>107818170
Thanks for the replies anons, but honestly I was hoping for something like "I have tried using two or more of CC/Cursor/Opencode/Crush/whatever to complete the same tasks from the same starting point with the same prompts, all using Opus 4.5, and found that X was the best".
That's probably not a fair ask, so I'll have to do it myself sometime soon. I'll post results if I do.

>Cursor has the right UX idea - AI first with the editor pane for reviewing
Actually I don't use it that way. My Cursor looks just like vscode with an extra AI sidebar on the right.
>dedicated TUIs wired to proper models
I only use Opus 4.5 in Cursor, so it's the same model as the CLI would be using.
>>
>>107818332
>but honestly I was hoping for something like "I have tried using two or more of CC/Cursor/Opencode/Crush/whatever to complete the same tasks from the same starting point with the same prompts, all using Opus 4.5, and found that X was the best".
I came really close to that with Claude Code, Cursor, and Crush, actually. Adding similar makeworkey tasks (wiring up various junk to make it "production ready on Kubernetes") to the same repo, all with Claude Opus 4.5 on the backend. The accuracy is the same for all chat modes, because it's the same model, on the backend. The difference is UX and how much of your local CPU it gnaws on doing the work. My ranking goes like this:

1. Crush
2. Claude Code
3. Cursor chat
4. Cursor tab-complete

I think that the tab complete is intended to be training wheels for people getting used to AI. Being able to go back and forth "yes like this not like that" with the machine helps it learn a lot faster than tabcomplete-and-fix-the-mistakes-by-hand, which in turn helps it generate better code on a first pass in the future.
>>
>>107818451
nah tab complete is great for small repetitive tasks. it's one of those things you struggle to live without after using it for so long
>>
>>107818451
Thanks, that's helpful.
I have a keybind set up to toggle tab complete and leave it off most of the time. When it's helpful it's very helpful, but find it far too distracting to leave on.
I try to avoid "back and forth" discussions with AI, because a long context full of things you don't want tends to give bad results. If its first attempt at implementing something isn't what I wanted, I throw it away and start over with a more specific prompt. If it's almost there, I sometimes add a follow-up but I think it's better to start a new context for the changes.
>helps it generate better code on a first pass in the future
This is only really true if you or the system updates AGENTS.md (or something similar), right? Otherwise each context is separate. Other than having more examples of correct code in the codebase for it to potentially find, I guess.
>>
>>107818789
>Other than having more examples of correct code in the codebase for it to potentially find, I guess.
Yeah that's the big one. Once the model gets a better sense of what the repo is trying to be and your naming conventions it's super smooth - just like mentoring a junior engineer!
>>
opencode is fine. The main downside is that it is built with JeetScript and takes up as much RAM as vscode. I used crush at first because it's just a single static binary, but it looks ugly.
>>
>>107812940
>>107812840
This is right out of the "sink your business" playbook we've seen others use. I worked for a company that did the same thing: Makerbot. They started off strong, had great support from the community, and were on a rocket ride to the moon. Then those morons decided to shut out the open source community and, overnight, they went from relevant to hated. Where are they now? A small arm of a larger corporation that has been subsidizing them for years. They lost community contribution and fell behind. Several layoffs, shipping jobs to China, and some shoddy launches later and they have become little more than a cautionary tale in some Github comment about pissing off the people that support you.

Like many others here, I canceled my Claude subscription. The cost to acquire me and many others was high. We recommend things to other developers (earlier this week I recommended Claude to two people, this is a typical week), and we make enough to pay for an expensive subscription. Some of us make purchasing decisions for companies. Retention was pretty simple and the customer acquisition was done. This decision will absolutely affect their bottom line and I won't be coming back. This weekend, I'll be testing other services and, after I find a new home to use with opencode, I won't turn back for a company that rug pulls like this.
>>
>>107817481
Whisperflow exists, it's probably not hard to integrate it into whatever assistant you want. The assistant is the hard part
>>
i use the cli mostly because my use case is to semantically search code (not edit it)
im not letting this shit touch code
>>
>>107818332
I haven't tried it on identical tasks but I've tried using cursor a few times, after getting used to claude code. My conclusion was that the prompts they use are different and basically I found it extremely inconvenient, but I believe that it might well be simply an issue of learning to work with a specific style.
What I found is that in cursor, the models were extremely eager to just go launch into implementing as much as possible, chaining together tasks and investigations and implementation etc. I'd ask a question along the lines of "do you think method X will work well here" and the model would go "great idea!" and immediately implement method X across five files without giving me a chance to give any further feedback.
With claude, the way I'm used to it is that I can hold a discussion, design things (even without plan mode or anything), and then tell it to implement and it will go off and do what we discussed, as we discussed. Some people call claude lazy but I haven't noticed that myself; its scope to me is great for what I'm used to, it actually completes the tasks I tell it to but doesn't take complete initiative and lets me keep a tight control on what it does and how.

So like I said that may be entirely just a difference in how used I am to the specific system prompt behaviours. I'm sure you can interrupt models in cursor to quickly give feedback for example, or pre-plan properly, etc. I'm just used to models having shorter run spans and a tighter feedback loop with me.

Also take this with a grain of salt because it was several months ago (I think I was still using sonnet 4.0 or maybe even 3.5), so models were noticeably weaker. I did try multiple models in Cursor, claude of course but also chatgpt and gemini o3-pro (I remember that was cool and new at the time) and found they all had similar behaviour, leading me to conclude it to be a consequence of the system prompt cursor uses.
>>
>>107819249
Genuine question but did anthropic change something? Based on the posts you quoted it sounds like opencode was using an unofficial workaround to basically access an unmetered API, and anthropic just changed their internal detection
It's just like youtube-dl which tries to download youtube videos without google's JS and ads, and has to use internal download methods and keep up when google changes those. That seems completely normal. Or more generally any kind of scraping is always going to be brittle.

As far as I can tell the public API is still perfectly functional and perfectly public in claude
>>
>>107819316
There's a mode dropdown just below the prompt textbox. It defaults to "agent" which is eager to start making changes, so if that's not what you want then you're expected to change it to either "plan" (writes a markdown todo list for you to review before pressing Build) or "ask" (read only, for research). I've set it to use plan by default.
I also have something in my AGENTS.md about doing the minimum that was requested, not making any related opportunistic changes, and not to bother with minor linter stuff.
>>
I really prefer IDE integration over standalone terminal tools, mainly because text editing is awkward in TUIs. The available operations for selecting/editing text are very limited and they're different from the ones I have muscle memory for. I'm used to jumping around with ctrl+arrows, then selecting some words with ctrl+shift+arrows, then typing to throw away and replace what I just selected. Or ctrl-backspace to instantly get rid of the last word I typed.
It makes it annoying to write and edit long detailed prompts, which is what you're doing most of the time you're interacting with these things. Maybe it's fine if you know vim and use its vim mode.
>>
>>107819337
they broke the workaround and now it's the top of hackernews daily complete with literal fedoras defending Anthropic while OpenAI just announced OpenCode support https://github.com/headllines/hackernews-daily/issues/2011
>>
>>107819528
>Maybe it's fine if you know vim and use its vim mode.
It is in fact fine. Trying to use CUA controls in a unix terminal is always a culture clash.
>>
>>107816038
Why use this over codex or claude code? If you already pay for one its free.
>>
>>107819869
Because it works with fucking everything else too.

https://github.com/charmbracelet/catwalk/tree/main/internal/providers/configs



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.