I like Claude, but wish they wouldn’t have restricted access to their own harness. I’m experimenting with OpenCode, Hermes Agent and Pi, trying to learn a good workflow with these tools. It’s interesting to try different hosted and local models. Not every problem needs a leading-edge model! But for coding elixir apps, Claude has a great out-of-the-box experience. IMO it’s important to keep looking at other solutions to avoid vendor lock-in.
AndyL
arcanemachine
Overall, it’s fine. As I’ve mentioned, the UI breaks fairly often. But my main pain point is this: The current UI paradigm great, but is probably not its final form, and Claude Code gives you very little in terms of being able to structure the interface between you, your machine, and the LLM.
Also, Claude Code is not open-source, and so its extensibility is limited to plugins, skills, etc. (This leaves you at the whim of a company that has proven to be very un-hacker-friendly.) So if something breaks (again, I’m affected by this fairly often), your only options are to roll back, or wait for them to fix it.
I’ll give you an example. (There are many, most of them are just papercuts. But this is just one of many.) Sometimes, I want to go down the rabbit hole in a conversational thread, but I don’t want it to pollute the main session context. But when I go to rewind the context, I want to summarize just the part that I’m rewinding. But with Claude Code, when you roll back, the “Summarize conversation” button summarizes the whole conversational thread, not just the part I rewound. Since Claude Code is completely un-hackable, I just have to talk to the agent and hope it summarizes the right part so I can copy it, then rewind, then paste the summary to the previous part of the conversation. (And don’t even get me started on the fact that the default option when you rewind the session is “Restore conversation and code”, meaning a single stray Enter key press can blow away your changes you made while on that conversational tangent.)
(Poor analogy incoming, just bear with me.) Basically, Claude Code is like Microsoft Word: It’s big, it’s got all the bells and whistles, but you better like what you get, because you can’t change it. Whereas I’m more of a Vim guy: I like to customize and mold the relationship between me and my data. So for me, the Pi coding agent is more to my tastes: Give me a basic, but extendable platform, and let me adjust it to my heart’s content. Considering how close the out-of-the-box Pi experience is to Claude Code, this doesn’t speak much to the bells and whistles offered by the latter.
Despite all this, I will say that Claude Code is an amazing tool. It’s just not the be-all-end-all of what a harness can be, and Anthropic further kneecaps it by keeping it closed source, and fighting the very community that makes it useful at every turn. Just today, they removed the ability to use claude -p without paying API prices. Thus further locking users into their walled garden.
I strongly encourage you to try it out. I have found GPT 5.5 to be quite capable (haven’t used it a whole lot, but have been impressed with what I’ve seen), and the lowest tier subscription gives you a decent amount of usage time. Plus, you can try it out with another harness! OpenCode is a great first step if you haven’t tried any alternatives. Pi coding agent is nice, but you have to put in the work to make it yours.
3
tacit7
I’ve been a user of claude max since last september. I just cancelled my sub this morning because of the drop of -p and sdk off the max plan. This will be my last month. Thank god Im getting a mac studio by the end of june.
2
mudasobwa
Creator of Cure
I am not going to pay a dime for Claude because Claude is far from being well-designed. My colleagues use Claude (the company pays for unlimited usage,) and I am paying for my assistant myself to avoid suffering and bleeding eyes.
This is not a place to advertise stuff (DM me to ask what am I using), but in my humble opinion the assistant in 2026 must have a comprehensive RAG, the ability to plug in my own RAG, the ability to learn from my code, and it must not need any sophisticated prompting/instructions to do their job.
For the good assistant I am ok to pay up to $2K in the foreseeable future (and I do pay about $1K now.)
2