Its useless now.. really weird.
What are people using? I want to be able to tell it "research x" and it should be taking its time, longer than 5 seconds, to come up with diverse sources and then start a really long long process to do that research. I used to have a bunch of skills that worked but things change. The defaults for ChatGPT SUUUUCK now sooo bad.
I’ve spent a lot of time on this and I got a feeling others might find it useful, so I am going to try getting it in the store (new to the store process) & I’d love some criticism!
(The 3 is a page I put up more recent to describe the agent control mode)
I made it for myself, I forget commands and so this has AI (BYOK right now) window that’ll stay on top and it’ll explain what’s going on in the terminal, generate commands, but its a lot more, most for saving time and some AI is purely for generating themes in the app (you can type stuff like “make the app look like Halloween and Christmas had a baby” it generates 5 quickly.. this was fun ti get working!)
It’s got personality and I went a bit too far with the themes and appearance stuff (can vibe code themes with AI or manual tweaks, vibe code shaders.. but this is just because I really like cool looking terminals😬🤷♂️ Buttons and a full user interface, nice developer fonts.
It has panels where you can set buttons to commands, all it does is paste it with an optional enter key. I have some remote Linux servers where I host websites and apps, I can just move faster when I don’t have to memorize key combos.
One of the best features (I think..I use it a lot) is the ability to click one button then paste it into any AI agent and it can control all the tabs, create tabs, use them at the same time, and you can see what it’s doing (it’s hard to see this otherwise since the trend is to hide most of it or just show you a summary.. Codex, Claude Code, any CLI agentic tool) it’s CLI with MCP as backup. Click one button then paste into agent window and it’s ready.
I don’t know what parts to monetize other than to have it be open source, and have an optional managed mode for the AI models so you don’t have to use API keys 🔑 I am super interested in experiences with the store and what types of things people might pay for or how to go about it (have it be free and gather data about usage than go from there?)
It’s got an experimental AI coding agent idle / stop detect & it’ll nudge it back to working again. It’s cross platform, rust, I really have NO clue what anyone else but myself would need in an app like this but it’s very stable and I use it everyday. I would add a feature when it was something I needed or wanted for myself.
Man.. both OpenAI and Anthropic do things that make it SO hard to keep a subscription- Chinese AI sites and apps work just fine they never have THESE types of issues…
If i am logged into the account that is the paid one it just DOESNT WORK. Sits there. Hours of googling and hacks, like deleting cache in browsers works for a minute, there is some other QUIC thing to disable which also makes it work temporarily but back to being shit and doesn’t work. How does it happen?
Like how ? How can a company be spending a trillion dollars and have such a shit website that fails so much for YEARS? How, I don’t see how science allows it. The laws of the universe literally say it can’t happen so what’s up? Why are there people 2 years ago complaining about same problem and it still exists and ONLY for paying customers not free? I tested that and it’s true. Free accounts don’t have the issue. Am I in a dream? I can’t get it to work there is no solution so is it done on purpose?
I got an even more annoying story about an issue with Claude Code that happens to tons of other people and just goes on for months and months or just forever. A problem so annoying, it’s literally death from blood pressure spikes unless you stop paying for their software. Who else has a story? Or explanation.
Cache deleting and disabling random chrome features has no effect this time so I just used Chinese GLM 5.whatever to do the task. Do they do this just to purposely make you not use more tokens since they’re loosing money? Why not just tell me lol I can’t not cancel a sub over this. It’s a SIMPLE CHAT SITE. They could fix it in 0.1 seconds but they aren’t so it’s gotta be done on purpose
Looks like there is some new competition for desktop apps.. has anyone tried it yet or analyzed it? I just found out about it
I am not sure how it’s still free but it is.. up to 12 new card generations a month (all the old ones will still work too)
Basically it hides your real debit card and makes a new one for every new transaction or you can set it so it’s one card tied to one store / seller.
Anytime you want to cancel a subscription you just open the app and delete the card, instead of worrying how long they will try to keep you on the phone or give you drama.
I also use it for when I wanna pay something that seems maybe shady but I’m not sure, so I just make a new card and set a max dollar amount so they can’t take more than some amount. Then I don’t gotta worry about scammers using my card.
It’s just such a necessity and there’s no downside to using it (been using for a few years now and no issues)
Using this link it says that we both get $5 I think it’s higher sometimes ($50)
I don’t even use it with Qwen models, I end up using it just when I want to use open source models but don’t have another harness in mind..
What harnesses are people using?
I use this everyday, its faster for me to look at a button, click button, it do stuff. I don't love pressing keys... nothing wrong with it, but I prefer to not type more than I have to.
Its the anti-key combo terminal app, no memorizing CTRL+ALT+WIN+F+50, ,. and no editing json config files just to change the color. Its got ALL the settings in a GUI. Its dripping with too much AI.
Its got AI help a click away when you forget a command, it'll generate it or do whatever you are trying to do. What do you think? Its got a sale already and i'm not totally done, but its in a totally functioning state / 99% good. Its free, and open source, I just have to change the repo to public but its on here:
Buttons CLI - Terminal app for people who hate memorizing key combos, hate "minimal terminal" apps (they all are, except mine!)
No memorizing key combos in this app! Press button, it do. Press button, it paste command. Press button, it do automation. Set up VPS servers.. just tell it to. Set up ai coding agents automatically.. just ask. https://buttonscli.com
https://poolside.ai/get-started
I have used its cli coding agent for a couple things.. and it seems good! It is updating docs while I do other stuff with Deepseek
curl -LsSf https://mistral.ai/vibe/install.sh | bash
People say Mistral sucks, but I am using it in one of my apps because it replies immediately, and so far, seems sufficient for doing things in a terminal, generating commands and automations and stuff like that.
https://build.nvidia.com/nvidia
Lots of free models, with rate limits. I haven't really 'tested' the rate limits so I don't know how reliable it is.
I did set this up which you can paste a OpenAI v1/models endpoint, optional key, and it will keep pinging it with requests (real requests, expecting to get output) and keeps a list of the models and whether they responded or not and date/time last checked. Its got some bugs so I wouldn't paste your real keys in it at the moment but i'll fix it in like a day or so / maybe today - the point of it was so i could go check URLs to see if a free LLM resource is even worth bothering with (its not, if models always don't work, or only work half the time) https://extra.wuu73.org/chu5
Opencode Zen and Go free models
I think you can use some without an API key.. not sure but if not just get an API key for Zen / Go (one key is good for both) and you can just use the free models without having to pay anything.
cheap:
Minimax M3/2.7/2.5 - Minimax's API / endpoints are reliable as f*** and last time I had a subscription, the lowest cost one was totally worth it and totally reliable, I could not reach any limits even using crazy amounts of subagents. Price went up I think.. I am considering trying the subscription again or just see if API calls will be cheaper...
Deepseek v4 right now.. tons of free and cheap flash, cheap pro, the Reasonix CLI agent seems actually pretty good (knock on wood)
StepFun StepFun Flash 3.7 is cheap, good with tool/agentic
Minimax highly recommended - assuming nothing changed with their extreme amounts of use they give you..
Opencode Go - I tried it, but ran out really too fast... like in a week, so I think it'd be cheaper just using raw API, and using their free models
I do keep a blog about all the cost hacks/cheap/free AI stuff here https://wuu73.org/aiguide - Its hard to find time to update it because things change so fast.
I started building ButtonsCLI because I wanted a real terminal with a non-minimal UI: preset buttons I can click, easier copy/paste, mouse-friendly text selection, and less of the "memorize a pile of key combos to do basic stuff" vibe. A lot of terminals seem weirdly proud of hiding useful UI, and I wanted the opposite. A more accessable terminal that doesn't put people off immediately.
I also wanted something AI coding agents could hook into without turning terminal work into a black box. The idea is that agents can control it, inspect it, and help with it, but I can still actually see what is going on. I don't have to start a fresh/blank session every time a coding agent needs to use the terminal.
It is at v0.7.0 right now: usable, free, and available on Windows and Linux (`.deb`, `.rpm`, `.zip`). A couple dozen people have already signed up on the website, which is enough for me to want real tester feedback before I spend even more time building things nobody ends up caring about.
If you like terminals but do not like memorizing everything, or you want way more control over how a terminal looks and behaves, I would love more alpha/beta testers.
buttonscli.com - lots of screenshots on there
## Some Features
- MCP and CLI control let AI coding agents attach to existing tabs/sessions, inspect what is already open, and interact with the terminal without forcing a brand-new session every time.
- There is a built-in `AI Help` window for asking questions, explaining terminal output, and generating commands, so you do not have to keep bouncing out to a browser tab. Full agentic mode is still a work in progress, but hooking up Claude Code / Codex / similar tools is already pretty straightforward.
- It has a lot of model/provider settings, supports OpenAI-compatible endpoints, and includes quick links/info for some free or cheap API options.
- The customization side is honestly overkill: 100+ built-in themes, a lot of fonts, separate styling zones across the app, shaders, GPU rendering, and AI-generated "vibe code" themes and shaders. Type "make terminal look like its halloween" will do just that.
- Preset buttons are one of the main reasons I made it in the first place. They are good for long commands, repeated commands, SSH logins, or anything else you get tired of retyping.
- There is also an experimental idle-detector mode that tries to nudge stalled coding agents back into working again. I have barely tested that one, so consider it real but still rough.
- There is a feedback button in the app for bug reports / complaints / ideas.
- It does not require an account or email to use. It does generate a random UUID for lightweight anonymous metrics so I can tell which parts people actually use.
- I still want to add optional encrypted sync for presets/themes/settings and maybe remote attach/control from a phone, but that part is not done yet.
I always have this issue every time I try to use Opencode, I thought maybe it just sucked and just quit using it since I never have this happen with any other coding agent.
So i'm wondering - surely someone has figured out a good way to just sort of smack it awake - i like it except for this one thing, i can't get anything done if i have to sit and babysit like this.
Like, is there a mode that I don't know about, that detects when there's an error and it retries/keeps going? I am seeing people praise Opencode all the time, everywhere, but this happens to me on Mac, Linux, WSL, it could be the model sure.. but all I need is something to see it stopped, and just give it a push to try going again. I will figure out a solution if no one else has one already but I don't want to reinvent the wheel.. so any ideas/solutions?