ZCode – Harness for GLM-5.2
zcode.z.ai
zcode.z.ai
I've never used IDEs and never will, why are these things being constantly shoved down our throats?
I also maintain https://github.com/nothingnesses/agent-images which allows you to use Nix to reproducibly spin up OCI container images containing agents and any other tools you need for development and use these with agent-box.
I use both at the moment to work on some personal projects with agents, where I set up multiple separate git worktrees for the agents to work in, preventing them from accessing anything outside of the worktrees and from trampling over each other's work.
What's your setup like and what do you use it for?
I have a M2 Max MBP with plenty of ram and I use VSCode + Zoo Code plugin with Qwen3-Coder-Next-GGUF:UD-Q4_K_XL to run local agentic coding sessions, but I'm intrigued by being able to run headless as I could probably run multiple instances in parallel to do stuff?
Like are you using UTM with some pre-built VM and a local LLM?
Curious.
shameless self-plug. I've been dogfooding it for the last 3 weeks now.
Edit: my theory is they wanted to mimic being the primary provider in a quick way with a lot of string replace. Though they could have added opencode back as a regular provider.
By the way, their repo was a bit weird with no changelogs at all. It seems to be picking up speed now with their communication. I actually read in the changelog just now that their Compose (plan/executre/review etc. something like that) flow is now deterministic with software instead of just prompts. That could be really good.
Harnesses are quickly becoming critical components of the "model" itself imo. Not shocking to me at all that a company that spots a revenue opportunity is keeping its harness closed source.
A harness is that covering every blind spot or sub-optimal but probable output people have hit in the wild, and a lot of problems just have better solutions if you say "break problem A into subproblem B and subproblem C, then solve".
Here on HN we discuss facts, jumping straight into racism has no place here.
That idea is wrong, though. These same people thinking harnesses are part of a moat are also boasting that s/ware is easily writable now.
There's no secret sauce in a harness that you can't vibe-code into your own harness.
I'm not saying you're wrong necessarily, but I do think that when the actions and words of a company conflict, it's a pretty safe bet that the words are just posturing and the actions better reflect their actual belief. In this case, regardless of what they're saying about software being easily writable now, they clearly seem to at least think there's something valuable in the harness if they're not open sourcing it.
Except you'd need the knowledge of what to vibe-code, no?
What knowledge? If you've used a harness, you know what it is supposed to do for you!
What further knowledge do you need that can't be extracted from an existing harness?
It is essentially a black box with full user permissions, meaning you are just handing over your entire system to a Chinese-owned server. With OpenCode and its GLM provider, at least I can monitor which files were read, which were edited, and what commands were executed.
Not to mention that Chinese national security laws legally obligate companies to cooperate with state intelligence and counter-espionage efforts [0]. If you have this installed on a corporate workstation, and your company is large enough, the possibility of them spying on you is not just a risk—it's almost a certainty.
[0]: https://en.wikipedia.org/wiki/National_Intelligence_Law_of_t...
no. they. are. not.
Some people are just terrible at it.
Dunno. I’ve moved to Pi for most of my work and it’s finally better than Claude Code for me; BUT the Anthropic models are definitely better in CC itself.
Also you don't need to believe me. There is enough evidence in the open source space.
If you're already used to your TUI coding agent, you don't need the desktop agent. Although it is nice that it is there for folks who prefer the Codex App/Claude App UI approach.
For some tasks, it's better. Opus refuses tasks for me pretty regularly. GLM 5.2 has never refused a task. So for anything security-related or that touches on topics that trigger Opus's safety guardrails, I use GLM 5.2.
OTOH, for anything related to UI design, I use Opus 4.8. It's much better at taking relatively vague descriptions of user interfaces and a mockup of a related UI and combining them into an immaculate design.
For anything else, I tend to run tasks in Opus and then have GLM review them and write a Markdown file with anything it finds. Then I have Opus review the markdown file and fix the issues it agrees with. The reason I usually go with Opus 4.8 first is mainly that it's faster. Opus 4.8 is, on average, about twice as fast as GLM 5.2 running on z'ai's infrastructure for the same task. There's a large variance (sometimes GLM 5.2 is pretty fast and Opus 4.8 is pretty slow), but on average it's a very noticeable difference.
When I run into Anthropic's Quota, I switch to GLM 5.2 rather than Sonnet. I don't think there's much reason to ever use Sonnet for anything if you can use GLM 5.2 instead.
This is all pretty subjective, of course. On average, I think Opus 4.8 is still a better, more reliable, and faster model, but if it went away tomorrow and I only had GLM 5.2, I wouldn't be too sad about it; I'd get things done with GLM 5.2 just fine.
One example I've hit is working on a benchmark of how well LLMs handle Kubernetes security tasks, there's a section on them exploiting security misconfigurations. Opus 4.6 was fine with that section, 4.7 and 4.8 saw some refusals and Fable point blank refused to do any of it.
The only other model I've seen refuse is OpenAI GPT-5.5, all the open weight models seem fine with it.
Ofc if you need to do that kind of work a lot you might be able to get on OpenAI/Anthropics allow-list for cyber work.
I've also had it refuse security-related tasks, and occasionally it stops without any discernible reason.
But when there was the Hantavirus thing a while back, I asked it if there was a vaccine under development and got a refusal immediately. I’ve had a few like that. It seems they’ve implemented really poor guardrails on certain topics (CBRN and cyber) that have lots of false positives. But if you actually chat with the model itself it’s quite lucid about what is legitimately dangerous and what is just performative “AI Safety” style refusal.
Edit- I see down-thread you use z.ai directly. Same concern, aren’t you worried about using it for professional stuff.
That said, GLM is worlds cheaper and a great little planner if you do a couple of rounds covering edge cases or things it might have forgotten. I can’t cache it via OpenCode Go, so I plan with GLM after gathering context cheaply, and then pass the plan to Gwen. That pairs well and shows the beauty of a multi-provider harness.
I should also add that I run GLM in Pi with a lot of cool little things such as hashline edits, some plugins for general smartness, etc. Opus runs in CC with the native tooling (except for Semble and Serena).
That said, it's interesting that they're releasing a bunch of stuff: ZCode, OCR.z.ai, Image.z.ai, Audio.z.ai, AutoClaw and some other stuff that https://chat.z.ai/ links to. That's a lot of stuff for one org to pull off.
Figured I'd try out their Pro coding plan, seems like it doesn't necessarily give me that much quota than Opus (at least given how many tokens are needed for accomplishing a certain task), but GLM 5.2 in of itself seems like a beefier Sonnet model, pretty good.
I guess the base is whatever the profit margin needs to be this month.
[1]: https://zcode.z.ai/en#:~:text=Base%20usage%20allowance%20inc...
[2]: https://support.google.com/gemini/answer/16275805?hl=en#:~:t...
Start plan: 5 Million tokens a day (GLM-5.2 3M, GLM-5 Turbo 2M)
For individuals: (+150% quota) $18.00USD+ For individual developers with a dedicated Coding Plan quota.
The app itself is interesting to me. I can see most of the agent trace (I can't see the tool definitions and the tool input args), I can set up skills and make the agent manage them and I can define sub-agents as well.
The UI itself is a bit weird, but I guess it's not thought to be a general purpose file editor.
Which make me raise a question. Why would I install a close source black box, that will send data to a country that you can't make legally liable for even most crazy miss doings.
The market of a hosted commercial version of glm is very weird. yeah you can deploy an open source version or run it locally, sure. This.... hm, i don't know why any company would take any risks to use GLM
Not to mention, claude and gpt only exist because of massive ip theft in the first place.
Lets move to caves and if I kill you first , i'm right, if not i lost. THat's your logic.
there is ToS, and every single person who is doing distillation knows it's illegal. Also on IP theft. There were several courts, in different countries they limit certain knowledge bases and thats it. so that entire claim is baseless.
Again this is the difference. A major book publisher brought ai companies to court and won. You can't doo anything with Chinese companies even if you are book publisher of this world.
1) If training a model on IP theft is stealing, distillation has the same ethical base and we can’t really complain there.
2) While Chinese companies have some protection, they still have to uphold the different laws to enter other markets. We can complain all we want about their immunity, but good luck buying a Chinese car that doesn’t respect EU-wide safety standards.
3) Most (all?) big US-based companies now put arbitrage into their ToS. So no, I can’t sue them either. At best, I can run an expensive uphill campaign against them, which will be mostly pointless in the end.
I'm wanting local context in the spirit of "here are 3 AI providers available, for coding tasks use this one... and for writing prose use this one... and for generating images use this one..." etc.
OpenCode was the first agent harness I used, and I have always like it. You can configure a wide variety of providers, but it's open source and has a number of core contributors.
The other opinionated option is Pi (the Pi agent harness). This is a great lightweight option and also supports a number of providers. You can also use local model servers.
i think people don't realize how much better life is over on this side, cc and codex rely entirely on vendor lock in imo.
I don't think I understand the token/cost implications of this feature
The orchestrator knows which AI client is running in any given worktree, so it would be fairly easy to designate which AI should receive what kind of tasks.
You run either Claude or Codex in tabs for each work tree. I do have some AI TUI specific instructions, for instance codex is primitive at monitoring compared to CC. So, there are additional notes for Codex workers on how to properly monitor for new "mail."
You work with the orchestrator on the primary worktree and allow it to delegates tasks to the workers and answer their smaller questions.
It surfaces results and assisting them with context clearing when needed.
The orchestrator and workers communicate using a simple shared file system under tmp/* and together they can handle a big and varied workload.
I use iterm2, so I’ve also added iterm2 specific python that allows the orchestrator to “kick” a worker or perform tasks otherwise veto'd by the TUIs (ie /clear) by modifying the input and submitting it.
I've been using it exclusively (and extending it, see https://a.l3x.in/ai) for months with mainly GLM-4.7 then 5.1 and now 5.2 and I could hardly be any happier.
I'm still working on a "Github/Forgejo first" based workflow but also quite happy with it already, basically most of my sessions run as a ci/cd job (triggered by "/pi" comments) and generate PRs or push commits to PRs, see https://github.com/shaftoe/pi-coding-agent-action
(Full disclosure: it’s my project)
It supports MCP (unlike Pi), sandboxing (with user-mode networking), and runs efficiently at huge contexts.
https://codeberg.org/mlow/lmcli
(The screenshot in the folder is a little bit out of date, but is still representative of the overall look)
I prefer having a GUI for diffs and session history,but if you prefer TUI you can just use OoenCode
Pi with some of the top plugins and hashline edits is very good. I stopped using MCPs because of their post about it, and instead rely on better command line tools.
Get the OpenCode Go subscription, add the plugin to enable Go’s caching, and my god, you’ve got nearly unlimited M3 calls. It’s not Opus-level quality, but that’s why you plan and review with better models. Open Router is also a good pairing for Pi.
To your specific request: sounds like you want custom agents :) That’s really easy to build in Pi, just ask it. My planning agent is basically a sub-agent-driven workflow that GLM orchestrates. You can have specific agents run specific models for specific tasks.
When looking at subscription offering by Anthropic and OpenAI, it's not even comparable, as a Codex $200 subscription can easily use a billion tokens per week on GPT 5.5 high/xhigh.
It's an interesting model from the perspective of being the most capable open weight model. But it doesn't have a solid place in this marketplace right now.
It shows potential, answer/code quality was solid, but I would need more time with it.
Its a hell of a lot cheaper though, so for me its worth it. I have more claude experience though, and I would say its almost en par with Opus 4.1
China have a history of stealing IPs/trade secrets and Chinese court favored its own local companies. while US have a robust court that can enforce IPs. if you want to risk your company's IPs/trade secrets/data for some cheap token. Go ahead and use Z.ai's services.
It is essentially a black box with full user permissions, meaning you are just handing over your entire system to a Chinese-owned server. With OpenCode and its GLM provider, at least I can monitor which files were read, which were edited, and what commands were executed.
Not to mention that Chinese national security laws legally obligate companies to cooperate with state intelligence and counter-espionage efforts [0]. If you have this installed on a corporate workstation, and your company is large enough, the possibility of them spying on you is not just a risk—it's almost a certainty.
[0]: https://en.wikipedia.org/wiki/National_Intelligence_Law_of_t...
I also consider Microsoft to be the biggest industrial spy in the world, them and google both are no doubt mining everything you type into office / gsuite, all your emails, etc. But nobody bats an eye when you write a word doc about some sensitive matter.
If my customers thought I was feeding their data into a Chinese owned LLM API (which to be clear I’m not), I don’t think it would go over well, and I’d be exposed legally to all sorts of things.
So the reason is risk aversion and desire to participate in US / western commerce. One can debate the actual threat, but why would you ever risk sending your data to a processor perceived as dodgy?
Suppose a US citizen, residing and working in the US and never traveling to China, crosses The Powers That Be. Which Power is more likely to do worse things to said citizen? Pretty unlikely they'll be rendered to one of the illegal Chinese jails in Brooklyn, more likely they'll be sent to Gitmo or a black site.
That's why, all other things equal, I try to keep my own government happy or ignorant, but don't really mind what I share with foreign governments, especially ones who won't forward the info to my own government.
Nevertheless, Americans thinking they are morally superior to China is always quite funny.
This administration is corrupt, cruel and doesn’t care about human rights.
And the worst is… Americans have voted for that administration…. twice!
I digress…
This has never happened in China, and will never happen, nor anything like it. Some open oversight is almost always better than possible secret oversight (and do you think that the Chinese government has user privacy on even its top 10 priorities?)
But yes, US intelligence has killed and ruined the lives of far more people than China has. Not sure how so many people buy into the narrative that they're protecting freedom and democracy.. They're protecting their freedom to kill and crush all their enemies and control every "democracy" on earth.
Reminds me a bit of the old “is your adversary Mossad or not Mossad” decision matrix https://www.usenix.org/system/files/1401_08-12_mickens.pdf
I'm no apologist for the US Intelligence and related organizations (not by a very long shot), but that is a very extreme statement to make.
Actually there are more such cases against the USA than China in public.
By the way, some pedant recently asked why anyone would run software with only a few stars. My thoughts on that are minimal: people can practice whatever slop logic they want. I've architected and built systems that handled tens of thousands of users. I'm not fucking around. The way I build isn't typical, and I don't suggest anyone try to mimic my approach, but it works for me and the way my mind processes complex systems.
To the peanut gallery: use it or don't, but don't give me a hard time unless you're ready to get one back. I've made plenty of mistakes in my career, and accountability is a crucial part of growth. I'm more than willing to work with anyone using my code, provided they bring valid, substantial criticism to the table.
The US is certainly inching in that direction but it’s not like someone from the US government sits at Anthropic’s HQ reading chats from state people of interest.
1) there is a very non-zero chance that the US government also has that data from OpenAI and possibly Anthropic
2) unless you are asking the chinese models to draw up plans to overthrow the chinese government, it's extremely unlikely they would ever care.
while china has a track record of harassing it's own dissident citizens abroad, if you're not chinese and not trying to subvert their government (or are a high-ranking government official yourself), it's kind of silly to suppose they would ever care about you or what you do.
and if you have information they want for their own national development purposes, like EUV engineers, they are much more likely to offer you fabulous amounts of money instead of try to intimidate or threaten it out of you.
PRISM ... XKeyscore ...
> The US is certainly inching in that direction
Itching to go in a direction that (publicly known) they have been in for decades now.
Do you really think the US government doesn't get access or couldn't get access to any of your chats with Claude?
You mean, like Windows and Android?
For GLM Coding Plan subscribers, quota consumed via Coding Plan for GLM-5.2 in ZCode is discounted by the coefficients below — the same usage draws down less quota, roughly 1.5x the effective allowance.
Peak hours (14:00–18:00 daily) 3x -> 2x
Off-peak (remaining 20 hours) 1x -> 0.67x
> Explanation and Recommendations Regarding Usage for Plan-Supported Models
> Note: Peak hours are from 14:00 to 18:00 daily (UTC+8).
Here's the message: "Cannot connect to API: write EPIPE"
(If this comment is too formal, I'm sorry. I used Google Translate to it [this line was NOT translated])
At 200k context that is only 85 requests for a whole week.
- https://igorwarzocha.github.io/howcode/
- https://github.com/ruuxi/stella
Not using Pi, but based on PI (no extensions possible)
But mostly vendor lock-in, I imagine.
It does have a 1.5x usage promotion for GLM 5.2 on the coding plan so now is a good time to test it...
But it already works really well with existing harnesses, I'm not sure why a dedicated one is needed?
I use it with https://swival.dev and everything works perfectly, no tool calling issues and it works fine with long sessions.
And most of the advancement and experimentation happens in some random 0-star github repos.
pi-tmux is one such example (seems to be archived now) which inspired me to use tmux as communication layer and provide visibility of subagents of multiple models in their native harnesses [1].
There's also herdr, which is not 0-stars, but is super interesting but lesser known project [2]. This also has interesting substrates to allow agent coordination.
None of these are harnesses per se, but they're pointing towards clear gaps in existing harnesses. For example, we've known for a while now that compounding knowledge of different class of models achieves better performance. Why is there no harness where this is a native functionality? And there's no harness where subagents are first class citizens both in terms of capabilities and UX.
Do they really correspond roughly? Seems like they’re trying to suggest a discount while still being worth a significant amount of monthly spend.
I've written up an explanation of what trips small models ups and how the harness can address that here https://yogthos.net/posts/2026-06-08-dirge-code.html
I suspect smaller models need more work than is practical to fit harnesses around. The smaller the model, the more work, and it doesn't carry over to other small models.
Deepseek r1 7b could not emit tool calls to save its life, gemma4 e4b couldn't get the names of files right, qwen3.5 4b gets stuck in dumb rabbit holes, I pointed it at a ruby script and asked it to run it, it tried running it with bash then got caught in a loop investigating.
Noble effort though! I guess I'll keep working on my barebones ruby_llm harness, with very tempered expectations. Each of these failure modes can be worked around, but there's too many of them to work around in the general sense.
I find I tend to view agentic coding similarly to a genetic algorithm. The model is the mutator function, and the harness along with the tests acts as the selection function. Each round the model generates some plausible code, it gets tested against the constraints, the model gets feedback and iterates on it until it converges on something that's workable. So, the real trick is to make sure the environment is producing correct pressures to guide the model in the needed direction.
Another interesting project in this space I can recommend checking out is ATLAS https://github.com/itigges22/ATLAS
For example, so far I haven't seen any evidence that LSP integration improves performance for small models vs using grep via a bash tool.
As it stands, it's kind of subjective, you just have to try the harness and see if the model seems to be have better than with the other ones you've been using.
Even without having hard numbers, it's pretty easy to see from the log whether the model is getting stuck or not.