Given how quickly things evolve, it’s easy to get lost in the numerous offerings and hard to get the best deal. So, what do you use? Both clients/harnesses and LLM providers or local setups would be interesting.
Personally, I’ve been using opencode with Github copilot for work. I’m currently looking for cost-effective provider for personal work. Maybe openrouter with one of the cheap models?
A mix of gpt-5.5 and claude opus 4.7, wrapped in my own mobile web UI that doubles as planning software and document editor.
The reason for the mobile focus is I’ve been trying to detach from my home office as much as possible, meaning I do a lot of up-front planning via ADRs and task/story writing, then I spawn agents to actually work on coding tasks. The only part that really requires sitting at a desk is code review, manual testing, and deeper human-in-the-loop coding or debugging sessions.
If you are able to max out the $100/mo plans from either OpenAI or Anthropic, then that is already a very cost-effective option.
I don’t like the term “AI coding”, but recently I have been playing with opencode with it’s hosted models and ollama for light local models. It’s good as a rubber ducky, an assistant, or co-programmer.
At work claude code, at home zed+deepseek. I loaded up a few dollars in deepseek’s API and they are lasting a long long time (many months). I also tried qwen locally but it’s too slow for me (it runs on the CPU…). I use AI though only for repetitive and boring implementation tasks, or throwaway one-time scripts that I don’t feel like coding by hand. The point of coding at home for me is to have fun, and having the AI solve all interesting problems for me defeats entirely the purpose of why I’m coding. At work we are unfortunately being pushed to use AI a lot more, and the amount of slop code is increasing :(
OpenCode Go/Zen have deepseek models that have the best cost-benefit out there. Then OpenRouter with selected providers if I want to try something else.
I don’t spend more than $20/month on personal projects and still manage to have a lot more functionality than I could have without AI.
Nothing. Fuck AI.
Claude Code, mostly, but I’m with Scipitie that the tool matters less than the process around it. What’s helped most is writing the project’s rules and conventions into files the agent reads each session, then putting the non-negotiable ones behind a linter or a test so it can’t quietly skip them. Treated that way it behaves a lot like the junior who’s read all the books and understood half of them. Left to its own judgement it drifts, which is the part the guardrails are there to catch.
Nothing. I threw them all out. After is forgot how to write good code and how to design good software.
That his how cerebral atrophy must feel like.
If you forgot to write good code and good architecture, in a span of what, at most, a year, then you were never a good engineer to begin with.
I was worth every penny I earned 4 years ago. 1 year ago, I had to open documentation for C++ features I use for a decade. I could not remember how it works.
Your brain is a muscle, and sure, I had more time to teach, but I got worse at teaching because I got worse at doing
Not remembering specifics of a technology, and completely forgetting the base building blocks - the same blocks that you should be using for AI generated code too, BECAUSE YOU NEED TO FUCKING REVIEW IT - is not the same.
I’m an Android engineer by trade. I might not be able to give you the exact interface definition of a BroadcastReceiver, or explain in technical terms the core differences between a TextureView and a SurfaceView (that’s what the documentation is for!), but for sure as hell can tell you if your architecture is good or not, or if the quality of the code you wrote is shite.
If I can’t remember a specific interface I can not judge if would be the right tool for the given job. Using suboptimal tools is bad architecture for me.
That’s just a simple example.
It took me about 8 weeks to get back up to my old productivity, and suddenly code that looked perfectly fine from agent backed project’s give me nightmares.
Not remembering an interface is easily alleviated.
Forgetting core architectural principles - which are the cornerstone of good architecture - cannot be fixed that easily.
Micro details, specifics, are what the docs are for. You don’t need to remember the specifics as long as you have the understanding of what the thing does.
Macro details - appropriate information and event pipelines, SOLID, KISS, etc., are what architecture is about. You can write the best micro-scope code if the end result on a macro level is spaghetti that would feed Rome for a year.
That’s not how a brain works.
People forget the language they grew up reading/writing/speaking simply because they stopped using it.
That’s… precisely how the brain works.
We’re not talking language specifics here. We’re talking about core principles of software engineering - principles that can be applied easily to other aspects of life. We’re talking about patterns, concepts, best practices.
Things that one needs to utilise daily even with AI generated code - that is, unless you’re checking in whatever the AI writes with zero review.
They actually do, and unless you talk in code this is more akin to a second or third language, not the one you grew up with. I have personally noticed how after a year or so of using the rng machine I simply could not code without AI, and am now lucky to be rehabilitating myself. There are already serious studies about the deskilling effect, this IS how it works. “AI” (meaning the LLMs now thrown at every thinking problem) is a lazy enshittifying tool.
I do not use AI to code; but I have been programming for two generations now.
I’ve taken a few breaks in that time, and switched technologies at least ten times also, and it’s very easy for me to forget a lot.
Programming is like any skill, it requires constant practice.
Am I a good programmer? Probably not that skilled, but I’ve written over a million lines of code and have run the gamut of architecture from: “Break this large file into 500 lines each, now we call that as a function” to “all numbers are object oriented” to “if it works great”
I use opencode with locally-hosted llama.cpp - usually with qwen3.6-35b-a3b.
I tried opencode go for a couple month, and its definitely nice to have an lln runner with more gram and more GPUs, but I prefer to have all my stuff local whenever it’s possible. Also, I’d use up my token allotments fairly quickly on opencode go.
I also tried opentouter and it, too, was great - many more models. But I exhausted by credits even quicker than opencode go, and its also not local.
What hardware do you use? How fast is it?
AMD Ryzen 9 9950X CPU and AMD radeon pro w7900 (48GB vram). I get 55tps output pretty consistently, but ingesting context starts around 1500tps and if context size reaches, say, 50K, tps drops to around 200tps. I often have to wait a bit, but it’s a price I’m happy to pay for local-only AI
Thank you!
Claude cli almost exclusively. So far mostly opus, but I need to start using less expensive models more frequently.
I have Codex installed, but I mostly use it to interface with Obsidian, which is where I track all of my tasks and organization stuff. Unfortunately, codex doesn’t play cleanly with Claude, so it’s hard to use it for coding at all, but I do have it review documentation and identify gaps or discrepancies, because it’s far cheaper than having Claude do it.
I’ve tried lmstudio for local AI with some ~30b models. It works fairly well and fast, but only for tiny context sizes. One I pass a few thousand tokens the speed falls off pretty quickly.
At work Cursor with Sonnet or Composer and very very few times Opus.
At home I can’t afford any of that so I use opencode and try to use super cheap models in openrouter and very few times to save money, I just delegate simple straightforward, tedious tasks to gemini flash or deepseek
I use Zed with either Sonnet or Gemini. I do plan to swap to using Qwen locally since my hardware has no issues running it.
I only use openwebui with an openai key that my employer pays for. No agentic stuff, haven’t found the need for that yet.
Good thinking to mention open-webui - I was only thinking agenetic coding, but I use open-webui for llm chatting. I think it’s fantastic.
I haven’t found anything that isn’t a shell backdoor that you let an llm run commands in, so strictly via chat interfaces.
I’ve been using oh-my-pi agent harness with a mixture of Claude for planning and either Gemma4 or Qwen3.6 for execution.
How do you get your LLM credits? Or do you run Gemma and Qwen locally? With which hardware?
For Claude I have the lowest tier subscription through work. I also have openrouter to use occasionally when I need it. Gemma and Qwen I run locally on a strix halo framework desktop I bought just before ram prices went to the moon.
. Gemma and Qwen I run locally on a strix halo framework desktop
So, CPU only?
No, Strix halo is AMD’s integrated CPU GPU using unified ram. Its the non apple tax equivalent of the mac minis people have been using to run local models. On one hand its a bit slower as it has lower memory bandwidth and that’s the limiting factor, but on the other its less than half the price for more memory and runs linux rather than osx.
Excellent, thanks for the explanation!
In work we use cursor. I generally use their composer model. Before that it was vscode with claude.
Personally i have not done any personal dev work in a long time, its mainly server maintenance and i can do that myself. When i get the itch to do more dev stuff ill probably use vscode with copilot. My personal machine isnt great for local models.
