I bought $18 GLM official subscription yesterday (5.2, but new model version was already leaking on some docs), set it up with Claude Code harness... and I’ve bumped to $80 plan almost immediately. It’s the first model that agreed on a proper security research (red team scenario), executed it seamlessly, including 0-days in WP plugins, RCE, 6.8 kernel exploit adaptation, etc - while playing against another GLM agent as a defender (following HF story)!
I understand that such models can be used by malicious actors, but it’s fair to have it publicly available (and play on your side in case of emergency). This is what changes the world in a better way, I think, not the guardrails.
At my work I have a $500 monthly AI budget. I have been using the $200 Claude subscription and most of my use is with Claude code. I think I'm going to switch to either kimi or glm and use the opencode harness. Both fable 5 and opus 5 have outright refused things like security related bug fixes and making monitoring tools. I am so happy that open models are good now
I generally use the $100-200 Codex/Claude subs, and have been blown away by the usage I get from OpenCode Go at $10/mo. At a minimum, excellent for automatically piping reviews to from Codex/Claude.
You should try a better harness. Try pi, or ohmypi if you want a good OOB experience
what is a harness? The comments below are mixing IDE/ADE but other suggestions are purely terminal things and I don't get what their value is over just a terminal. Is a harness like a loop where it's just a vague thing that everyone nods about but everyone is nodding at something different?
My understanding is that the harness is the set of function calls (or tool calls) that let the LLM interact with your codebase. It's independent of the IDE or CLI.
The tool calls will be, among other things, something like ReadFile, RipGrep, PatchFile, Shell.
When people talk about the value of different harnesses, they're also implicitly talking about the quality of the system prompt.
The same exact model, when given a different set of tools and a different system prompt, can behave differently.
harnesss == thing that calls LLM API, acts on response, and maybe does that again.
You sound like I could afford that.
I'm gonna shamelessly plug my own here :) https://dirge-code.github.io/
I like it.
I think creating your own agent is the Hello World of agentic coding. Instead of Rust, I used D for mine.
That actually looks nice
I’m in the Claude code harness for everything boat too. What are the alternatives?
What the person above is suggesting:
* https://pi.dev/
* https://omp.sh/
(no personal opinions of either, links might be useful)
I think that OpenCode is nice, their CLI version is enjoyable and their desktop/web version is okay:
* https://opencode.ai/
I also quite like driving OpenCode through something like Kepler / Paseo and tools like that (with those I can still use my Anthropic Condition by Claude Code being treated similarly - as something that gets tasks dispatched to it, while the GUI I see is Kepler / Paseo).
On the desktop side, ZCode was surprisingly usable for something that came out of nowhere (I wasn't aware of it at all before trying out the GLM Coding Plan): https://zcode.z.ai/en
Last time I tried some of these, none of them had the "manual mode" that CC has, where it shows you change by change as diffs and you can edit them before accepting and moving on to the next change. I like that because if it's going off pattern I can spot it early on and guide it correctly, instead of having to review the whole completed diff at the end when it's too late. I should spend the weekend checking them out again to see if they added that but I assume with everyone going full agent mode they probably didn't.
The philosophy with Pi is it is minimal (but functional) out of the box and easily extensible. I'm not familiar with that feature but I would not at all be surprised if someone already coded a Pi extension that does it.
Both Pi and OpenCode let you customize them. You tell the AI you want "something like claude code manual mode", and they'll modify your configs to do the same thing, or build an extension for you
(however, it's much faster to use Plan Mode to build a plan of what it will do, and then execute the plan in Build Mode. you can also have the AI make a script that will be executed deterministically)
No one in their right mind would install software using `curl | bash`
I feel the same, which is why I only use these tools in a docker container. Because life is compromise.
T3 Code has been amazing. Completely free. Really impressed with the desktop app and the mobile app experience and the way it works seamlessly has me actually accomplishing tons of stuff while I'm out on mobile that I would otherwise have to wait to come home for. First time in a while I'm actually excited to use a desktop UI instead of the terminal. Blows away the official Claude Code mobile app. I can switch between my Claude and Codex monthly subscriptions in it as well. There's a TestFlight beta SwiftUI mobile version that's so much nicer than the one in the App Store. I'm running the nightly version of the desktop app.
And this is coming from someone that's not particularly a big fan of Theo. T3 Code should get more recognition; people aren't just aware of it yet.
ArtificialAnalysis puts out benchmarks for harnesses now as well, and OpenCode seems to be winning it. https://artificialanalysis.ai/agents/coding-agents#coding-ag...
I only found this yesterday, and it inspired me to start testing out OpenCode.
Just a warning, this is on Opus. There's not a clear harness winner. It will change depending on the models.
I'm not surprised to also see Cursor above Claude code, their harness is very good.
In what scenarios?
It indexes the code efficiently, seems to find stuff quicker, it has a very nice UI (much better than Claude Codes IMO), it has a nice sub-agent UX which I find triggers more reliably, diffs render nicely. Otherwise it just seems to work in a purely vibes sense.
That said Claude Code is perfectly fine. I just prefer the integrated experience of using Cursors since I already use VSCode, but I still mostly use Claude Code because of their Max/Fable plan.
Can OpenCode dispatch background subagents yet? I tried it a week ago and saw nothing. This is 99% of my workflow at this point.
Having just "asked" my opencode instance-- Yes, but it's behind an experimental flag and not necessarily feature-complete.
In the new v2 beta, yes. Major QoL upgrade, so much less sitting around waiting.
The v2 branch of OpenCode has not been touched for months, if it's beeing developed then I don't know where.
this is very easy to check. i promise i'm not tricking you and just uploaded this. https://github.com/anomalyco/opencode/tree/v2
this is the integration branch for https://opencode.ai/v2 . it has been for months. it's where the Effect-based refactor has been landing.
[dead]
It’s not just bloat at this point. I run oMLX and run models locally. using Claude code on the first message dumps 40k of tokens that my laptop takes 5 mins to compute.
I’ve stopped using it completely now.
If you like running everything in a VM and using a web browser as your UI, Shelley is very good: https://github.com/boldsoftware/shelley
It works nicely in the browsers on my tablet and phone, too.
On exe.dev you can ask it to customize itself, and it will automatically rebase your customizations when upgrading to a new release.
https://github.com/tontinton/maki is tackling the right issues IMO. not sure how they compare with the rest
Piggybacking on this thread to ask my question: What are alternatives that are multiplayer (team oriented) by default? For example, I want my team to see all my sessions easily, vise versa. another way of stating: all the agents are running in a container that that any member of the team can view and interact with.
Mine is a WIP for automation but has similar concepts to what you're looking for: https://github.com/rush86999/atom
I think that writing your own harness is a rite of passage now, just like writing your own search engine or database, rolling your own crypto…
Anyways, please try mine!
https://github.com/deepclause/deepclause-sdk
thank you all! got something to tinker with this weekend
i like to challenge my assumptions and try new tools
Just as a +1 anecdote. I enjoy using pi a lot. I used to h think the harness matters a lot but with the current iteration of models I am starting to sway that while it matters it’s less and less important and that CC is bloated. I did some quick tests when I switched and a task that would take $5 in tokens would be completed in $0.50 in pi. Very anecdotal and I don’t have a test framework setup to make this very official but increasingly felt like CC was spinning its wheels on the easiest of tasks.
are ohmypi and pi related?
that's a very compelling use case, thank you
[dead]
I’ve tried a bunch of them, and I seriously do not understand these recommendations. It was a rough road and a steep hill, but right now CC is absolutely the best harness on the market, as for me, whatever top tier model is under the hood (mostly, some of them, like DeepSeek, don’t fit CC at all).
Inversely I don’t understand the praise for CC. These days it feels like bloatware. It absolutely can get the work done but when I measure on token and time use it ends up being a multiple of pi like harnesses.
CC works but for me it felt like increasingly they have zero incentive to make it a great experience. You hear folks like Boris talk about spinning up thousands of agents over night and agents chatting back and forth in GitHub issues and while I think it’s great from figuring out what the future looks like I don’t think it represents the reality of ROI today. So the folks building the tool are so disconnected I am simply not sure it’s a great experience anymore.
So is the quantitative difference in token use the only difference or do you think there's also a different qualitat? I'm on CC only and immensely happy. Very productive both at work and privately and at work I average around $250 a month which probably means nothing but it's little compared to my salary.
Is that the main concern though, cost?
For me, at least it's that the newer Claude models seem optimised for one-shotting things, which is not what I want. As the amount of code per turn increases, I have a harder job keeping up and ensuring that it's doing what I want.
That being said, I had to nope out of a similar thing from GPT 5.6 today, so it appears to be a US frontier lab issue. Claude is particularly bad though, as it produces far too much code even when I tell it not to, unlike GPT (and Kimi) which at least listen to me a little better.
More generally, I want a usable human review experience, and Claude code doesn't deliver that for me.
Quality is hard to measure and I would not say the concern is so much cost but the intersection of cost and time. Often I am jamming on something and I like being somewhat in the loop. So maybe same level of quality, I am using Anthropic modela for both harnesses, but I get to the output quicker and at a drastically lower cost.
Funnily enough, I would say almost the opposite. CC’s feature set is basically table stakes for an agent these days (does it have ACP yet? Very close to behind table stakes if not) and it has a lot of bloat powering that.
IMO part of it is that the underlying LLMs have gotten better enough that harnesses feel better even if they haven’t changed. I have a toy harness that barely implements the features you’d expect and it works surprisingly well. Like there’s literally nothing clever, it calls tools and that’s about it, and it still mostly does the right thing.
Why would anyone ever need ACP? I'm not trying to be an asshole. I just seriously don't understand the value proposition.
Edit: lol, I don't think ACP is even actively developed anymore. It seems to have been merged into another seemingly pointless standard with an even worse name, A2A. [0]
[0] https://github.com/orgs/i-am-bee/discussions/5
It's buggier for me than it has ever been before. I don't think that agentic coding always leads to such a buggy mess. I just don't think that the Anthropic front-end software team is very good at agentic coding.
what is this comment based on ? vibes?
https://artificialanalysis.ai/agents/coding-agents#coding-ag...
Based on the fact that Claude Code is only optimized for Anthropic models, whereas Pi and Omp are optimized for a wide variety of models, including open weights.
they are not really optimized for 'wide variety of models' . what optimization did pi do for glm 5.3?
Tool calling success rate, in the case of omp.
because not everything is a shilling advertisement?
Vibes like your low quality comment?
What’s the counter argument? pi and ohmypi are pretty fantastic. Of course like all developer tools it depends how you do your work but I am not sure what you are trying to achieve in your comment.
how would i comeup with counter argument if i dont know what original argument is. No one is disagreeing with your subjective experience, gp comment said 'better' without qualification.
So you’re just vibing. Got it.
“ Open Source: We will release the weights in two weeks after launch, once safety evaluation and hardening are complete.”
Cybersecurity capability might be nerfed
We'll unnerf them, don'r worry.
Yes, I am tired of Claude and GPTs. I am ready to diversify my $300 per month on other vendors. Will try GLM. How was your rate limits and availability experience on $80 dollar plan?
It’s comparable to Anthropic usage, to be honest. 2x GLM agents ate 18% of weekly usage on this mid-tier plan within ~8 hrs (non-stop work, a lot of tool calls, appx 4 compactions each), I think. I didn’t make a proper statistics snapshot, sorry.
> I understand that such models can be used by malicious actors, but it’s fair to have it publicly available
I feel like there should be some mechanism to prove you own the code/app/site/whatever and it will remove the guardrails from the LLMs allowing them to find and fix these vulnerabilities.
This is a "they have guns so we need guns" scenario.
You can't guarantee everyone else will use a neutered model.
Isn’t this essentially what anthropic is doing, albeit in a manual fashion? They work with code owners to run mythos and find issues.
Only if you're some big corporation with deep pockets. They actually accepted me into their cyber program but Fable's still locked down.
OpenAI now makes it easy to join their verified security program. Took me 5 minutes, and I was able to get GPT to do a full end-to-end pen test
Impossible with source code, possible to bypass with app/site
Don't we already do this with services like Let's Encrypt, which is arguably more sensitive? If you had the codebase you could fake it, but it would still provide some amount of protection against abuse.
With Let's Encrypt, all the verification is done on their side with them controlling the connection between themselves and whatever they're trying to verify.
In this case, you can put whatever you want between the harness you're running (or modify the harness itself), and essentially "lie" to the model. Any verification technique would be fairly trivial to bypass, while you continue to run the harness locally.
How much usage do you get out of it per week? How many millions of tokens?
Anthropic was stingy as hell with its Fable and cybersecurity nonsense, switched to OpenAI which is much better but still not enough. I'm tempted to switch again...
On $80 plan 200M tokens ~20% of weekly usage
[dead]
They're invaluable for developers to fix their code. This is definitely an area where AI decisively beats human devs in a very valuable way. It can try so much surface area so fast.
If it won't attack my stuff, it won't help me build my stuff to be secure.
Exactly my CoT! I hope z.ai won’t change this behavior after training it on our input the same way as Anthropic did (shame on you, folks, seriously)
> after training it on our input the same way as Anthropic did (shame on you, folks, seriously)
What do you mean with this? Honest question!
Fable wasn’t trained on github commits, let’s be fair.
how do you configure claude code to use GLM ?
https://docs.z.ai/devpack/latest-model#switching-models-in-c...