HACKER Q&A
📣 _345

What open source LLM agent do you guys use and why?


I have tried aider, aider-ce, and opencode. I'm on Windows 11, I code in Windows using MSYS2's bash-like shell and in Ubuntu via WSL2 (so just bash).

In my experience, aider is the best out of the three. It has the least features too, but it "just works" the best. I tell the tool exactly what I want done and after learning a bit it goes and does that right like 90% of the time.

With aider-ce and opencode, I ran into way too many bugs. For example in Windows right now it seems that copy pasting just doesn't work, which is a massive loss of functionality when you can't copy paste important snippets to the model.

Opencode looks really cool as a TUI with the themes and such, but I really didn't like how random these agents are. I don't know if claude code and codex are like this too, but whatever simple task I give out could be solved right away or it could for some reason result in taking 5x as long as aider as it calls random tools, tries to do web searches for documentation that is technically correct but not the thing it needs, running into 403's probably because of bot detection, and then trying that 3 more times before it submits a worse patch than what aider would have made.

I'd just stick to aider, but aider-ce (community expansion) exists because the dev just kinda stopped working on it despite the fact that it has as many stars as opencode. Wondering if there are any better alternatives out there or maybe if others feel the same way about what's happening to aider. Its a shame because aider really is great at just applying AI just enough that I don't have to stress about its stochastic nature and it just does exactly what I asked no more no less. I feel like I have maximal control over AI running on my code with aider.


  👤 mindcrime Accepted Answer ✓
I tried OpenCode and want to like it, but there were major issues[1] running it in Konsole (and most other Linux terminals that I tried). So I gave up on that for now. I spent a little bit of time evaluating Codex and that seems OK. But so far, to the extent that I use any "coding agent" it's usually GHCP in Agentic mode. But truth be told, I still mostly code directly and just use Copilot for automated completions. Maybe eventually Codex or one of these other agents will win me over. I'll be happy to try OpenCode again if they fix the terminal issues.

[1]: https://github.com/sst/opencode/issues/2697