# ZCode: a desktop alternative to Claude Code on GLM-5.2
> Installing ZCode from zcode.z.ai, five days of access without a card, GLM-5.2 plans, comparison with Claude, and data security limitations.
Source: https://ailoot.ai/blog/en/zcode-claude-code-alternative/
Canonical: https://ailoot.ai/blog/en/zcode-claude-code-alternative/
Language: en
Publisher: Ailoot (https://ailoot.ai/)
Translation: guide-240
Revision: sha256-f85bfb084878f541e294b48c025a3ba5d7e4a2dcd6494e4fa73eb1f45d47c36a
Key takeaways
On July 2, Z.ai (formerly Zhipu) released ZCode, a free desktop coding application powered by its GLM-5.2 model. In essence, it is Claude Code, but in a separate window rather than a terminal, with files, Git, and a website preview, paired with GLM’s lower API token rates. The cost of a successful task still depends on usage and result quality. Below, I explain where ZCode shines, how far GLM-5.2 trails Claude in benchmarks, and whether you should hand your code to a Chinese service.
Learn about working with the service in the [Ailoot documentation](https://ailoot.ai/docs/).
## What is ZCode, and why is everyone talking about it?
Key takeaways
ZCode is a standalone desktop coding application released by the Chinese company Z.ai (formerly called Zhipu). Inside is an AI agent that reads your project and writes and edits files on its own. It runs on the company's own GLM-5.2 model. The main draw is simple: it does what Claude Code does, but in a convenient window and at a noticeably lower cost.
I have long said that whenever a functional, inexpensive alternative to Claude Code appears, it is worth talking about. Back in early June, I [pitted Claude against a Chinese model](https://ailoot.ai/blog/en/glm-5-2-vs-claude-code/) and said plainly: if you cannot use Claude, the Chinese alternative looks very good. Now Z.ai has gone further and built a whole application around its model.
ZCode came out on July 2, 2026. It is a program you install on your computer, like a regular editor. Open a project folder, describe a task in words, and the agent figures things out on its own: it reads files, writes code, and runs commands. It is exactly the same idea as Claude Code and [OpenAI Codex](https://ailoot.ai/blog/en/migrate-claude-code-to-codex/). The difference comes down to two things—the interface and the price—and that is what this review is about.
Why all the buzz? There are three reasons. First, eight updates came out in less than a week around ZCode; the tool is being refined literally every day. Second, it is powered by the new GLM-5.2 model, which is close on public benchmarks to top-tier Claude, while costing several times less. Third, Z.ai offers five days of free access without requiring a card, so anyone can try it. For an entrepreneur, that means one thing: there is another serious option on the table, and you need to figure out whether it is right for you.
Let me separate two questions right away so they don't get mixed up. GLM-5.2 is the model itself, the “brain.” ZCode is the application, the “hands and window” around it. I have a [separate review](https://ailoot.ai/blog/en/glm-5-2-vs-claude-code/) of how close the model itself has come to Claude. This article is about the application: how it works, how it is more convenient than the terminal, and whether you should trust it.
## How is ZCode different from Claude Code?
Key takeaways
The main difference is obvious right away: Claude Code runs in the terminal, while ZCode is a full-fledged window with files, a terminal, Git history, and a live website preview all side by side. For anyone intimidated by a black command-line screen, this is a whole different level of accessibility. Under the hood, both do the same things: show a plan, edit files, and run commands.
Open ZCode and you see a familiar application window. Project files are on the left, the conversation with the agent is in the center, the terminal is at the bottom, and there is a Git panel and a live preview of your website nearby. Everything is in one place. By default, Claude Code is a command line: powerful, but bare-bones, and for a beginner, a black screen full of text can be more off-putting than an experienced person might think.
I know this from experience. When I was putting together my internal work interface, I wanted a window where I could see the process—as an entrepreneur, that feels calmer and clearer than following lines scrolling by in a terminal. ZCode addresses exactly that need: it shows you what is happening instead of making you keep everything in your head.
Now for the features. ZCode and Claude Code are more alike here than you might expect:
- **Plan mode.** ZCode can first show a step-by-step plan and wait for your approval before touching the code. Claude Code can do the same—I covered its [goal mode and planning](https://ailoot.ai/blog/en/goal-mode-claude-code-codex/) separately.
- **Goal mode.** Type `/goal` followed by a task in words, and the agent runs in a loop, checking after every pass whether it has reached the goal. It keeps going until it succeeds. Handy for tedious tasks: fix all the type errors, get the tests passing again.
- **Agent autonomy levels.** Five modes, from “ask before every edit” to “do everything yourself; don't interrupt me.” Switch between them with one keyboard shortcut. And this is no small thing: it is how you balance speed and control.
- **Project rules file.** ZCode reads `AGENTS.md`, a text file where you put your project rules. Claude Code uses `CLAUDE.md`. The formats differ, but the idea is the same, and [you can use one file for both tools](https://ailoot.ai/blog/en/agents-md-claude-md-one-instruction-file/).
ZCode is not limited to GLM. In settings, you can connect any model using your own key: Claude, GPT, Kimi, and others. That means you can keep the ZCode window as the interface and put Claude under the hood. The application and the model are two separate things, which gives you a freedom Claude Code does not: there, the model can only come from Anthropic.
Let's put the differences in a table:
| What we're comparing | ZCode | Claude Code |
|----|----|----|
| Interface | Separate window with files, Git, and preview | Command line (terminal) |
| Learning curve | Lower—everything is visible | Higher—you have to get used to the terminal |
| Model under the hood | GLM-5.2 by default; any model can be connected with a key | Claude only |
| Project rules file | `AGENTS.md` | `CLAUDE.md` |
| Goal and plan modes | Yes | Yes |
| Age | Released July 2, 2026 | On the market for over a year; mature |
## How good is GLM-5.2 at coding in practice?
Key takeaways
For everyday tasks, it is indistinguishably good; on the hardest ones, Claude is still ahead. On a demanding test of real tasks from projects, GLM-5.2 scores about 62 out of 100, compared with 69 for top-tier Claude—a gap exists, but it is small. For most vibe-coding projects, this is more than enough.
It's easy to get taken in by impressive claims, so let's look at the numbers. There are benchmarks that give agents real tasks from live projects and measure how many they can solve on their own, without human edits.
| Model | Test of difficult tasks (SWE-bench Pro) | Comment |
|----|----|----|
| Claude Opus 4.8 (top-tier) | about 69 out of 100 | the market's upper limit |
| GLM-5.2 | about 62 out of 100 | 7 points behind |
| GPT-5.5 | about 59 out of 100 | for comparison |
Source: [Z.ai’s official GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2), which reports SWE-bench Pro scores of 62.1, 69.2 and 58.6 for GLM-5.2, Claude Opus 4.8 and GPT-5.5 respectively. These are provider-reported results; its footnotes describe the evaluation harness and settings. They are not a measurement of your project or proof that every task will behave the same way.
A 7-point gap is a real difference, but a small one. On the trickiest tasks, where you have to overhaul a large project from end to end, Claude is noticeably more reliable: in a separate test, it solves nearly 89 tasks out of 100. But vibe coders rarely face tasks like those. In everyday work—building a website, editing a form, connecting payments—the difference between 62 and 69 is barely noticeable.
Put in terms an entrepreneur can understand: you won't notice a difference between GLM and Claude on 95% of your tasks. You will notice it only when a project is large and complicated, and at first, you will hardly have any of those. So, to answer the question “Is GLM good enough at coding?” briefly: yes, it is.
The model itself is only half the story. I have seen vibe coding fail many times for one reason: the agent was not given context. Give an average model the right setup—the project structure, rules, and a second brain—and the result will work. Without that setup, even a top-tier model struggles.
## How much does it cost, and is it really five times cheaper?
Key takeaways
The application itself is free; model use is paid through the selected API or Coding Plan after any applicable trial. Check the provider’s current model quota and expiry. The launch-era subscription figures below provide context, while the API comparison uses token rates; neither proves a fixed fivefold saving for every completed task.
I'll start with the good news: ZCode itself costs nothing. Download it, install it, and use it. You pay only for the model you connect. Many editor agents work the same way, and the guide to a [free AI agent for VS Code](https://ailoot.ai/blog/en/free-ai-agent-vs-code/) covers ways to avoid paying even for the model.
Trial quotas are measured in tokens, not words, and vary by provider and model. The [official connection guide](https://zcode.z.ai/en/docs/configuration) currently describes a five-day BigModel trial with 8 million tokens per day split between GLM-5.3 (3 million) and GLM-5.3-Flash (5 million); it also describes trial access through Z.ai. Check the selected provider’s live model buckets, remaining quota and expiry in ZCode rather than treating the older GLM-5.2 allowance as a permanent offer.
After that, there are two options. The first is a Z.ai subscription for its model:
| Plan | Monthly price | Who it's for |
|----|----|----|
| Lite | $18 (currently about $12.60 with a 30% discount) | trying it out; light tasks |
| Pro | $72 | intensive daily work |
| Max | $160 | many tasks; demanding projects |
The second option is to pay directly for API usage, by token count, without a subscription. The [official API price table](https://docs.z.ai/guides/overview/pricing) lists GLM-5.2 at $1.40 per million input tokens and $4.40 per million output tokens. Those are token rates, not a measured price for one task: compare actual input, output, cache usage and retries before claiming a fixed saving against another model.
One caveat about subscriptions. If you compare monthly plans, the difference is milder: Claude's basic plan also costs about $20. API token rates and subscription allowances are different comparisons. Neither establishes a universal fivefold saving for the same successful task. The lower API rates make GLM worth testing on a small task, but record the complete bill and verify the result before deciding whether it is cheaper for your own work.
## How to install ZCode in 10 minutes
Key takeaways
It installs like a regular program: download a file from the website, drag it into Applications, and launch it. Next, click “Connect,” choose a model (GLM or any model using your own key), open a project folder, and you can start assigning tasks. The whole process, from download to your first command, takes about ten minutes.
Let's go through the steps, without technical fog.
1. Go to zcode.z.ai and download the version for your system:
- Mac—the DMG file; open it and drag ZCode into Applications.
- Windows—a regular installer; run it and click “Next.”
- Linux—still in testing, but a version is available.
2. Open ZCode and click “Connect.”
3. Choose a model:
- For GLM, select the model currently offered by the provider and check its trial eligibility, quota and expiry before starting; the available trial buckets can change.
- If you want Claude or another model, enter your key for it.
4. Open your project folder directly in ZCode.
5. Describe your task in words or use the `/goal` command, then watch the agent work.
One tip from experience: don't put the agent in “do everything yourself” mode right away. Start with a mode where it asks before every edit—that way, you can see its reasoning and catch mistakes before they get into the project. Give it more freedom gradually, once trust has been established.
And keep in mind that free access still has a limit, even if it is generous. Run the agent on a real task, not an empty trial—five days is enough time to form your own opinion.
Before connecting any service to a work project, ask yourself one question: what would happen if data from this project ended up with a third party? The answer matters more than the difference in benchmark results between the models. The next two sections are about this.
## How can you check ZCode's access and payment terms?
Open access does not guarantee convenient payment: payment terms for services like these change, so it is best to check on the spot how to pay before subscribing. Even without a direct subscription, GLM-5.2 is available through intermediaries; the model itself is open.
## Should you trust a Chinese service with your code?
Key takeaways
There is a big “but” with ZCode that means you should not recommend it blindly. Using Z.ai-hosted GLM sends model requests to that provider; using another model requires a separate check of the application’s routing and telemetry. Provider data handling and applicable law matter for sensitive code. This is written into law, and there has already been a case where Zhipu was caught collecting more data than it should. The conclusion is simple: for personal projects and trial runs, it is fine; for sensitive client code, think twice.
I don't like to scare people, but this is something I can't stay silent about. When you use GLM through Z.ai's servers, your code, your requests, and what the agent generates all go to China. China also has its 2017 National Intelligence Law: Chinese companies are required to cooperate with intelligence services and hand over data when asked, regardless of where their servers are physically located or what their privacy policy says.
These are not guesses. In May 2025, a Chinese cybersecurity center itself flagged the Zhipu app for collecting more data than the user had authorized. So the issue is not only the law, but also something that has already happened.
According to Z.ai's own documents, the following information is sent to its side: account data, technical information about your device, your requests, generated code, and records of conversations with the agent. That is not a problem for a side project or a trial. But if you are writing code that contains customer data, passwords, or trade secrets, you should think carefully.
This is exactly the topic I talk about all the time: [vibe-code security](https://ailoot.ai/blog/en/claude-code-data-access-safety/) is a hygiene practice worth getting used to. Here is what you can do if you want to use GLM but are worried about your data:
- **Don't feed it secrets.** Keys, passwords, and customer data should not end up in a project you open in ZCode. This rule applies to any cloud agent, not just a Chinese one.
- **Separate your projects.** Use GLM for personal and trial projects, where it is cheap. Use a tool you trust more for work involving sensitive data.
- **Connect your own model.** ZCode can work with Claude or another model using your key. The window stays and the selected model endpoint changes, but key ownership alone does not prove where the application proxies requests or sends telemetry. Verify routing and privacy behavior for that mode before sending sensitive code; the connected provider’s prices also replace GLM’s rates.
Running open GLM-5.2 on your own hardware can remove the need to send model requests to Z.ai’s cloud, but it does not by itself prove that the surrounding application has no telemetry or other outbound traffic. But that requires so much graphics-card memory that it is not a realistic option for an ordinary person. So here is the honest conclusion: GLM is great for light tasks; think three times before using it for secret code.
## So, should you switch to ZCode?
Key takeaways
If Claude is available to you and you are happy with it, there is no need to drop it; GLM is still a little weaker at complex coding. But ZCode is worth installing and trying in two situations: if Claude is unavailable or expensive for you, and if you are more comfortable working in a window than in a terminal. Five free days should be enough to decide for yourself.
Let's give a direct answer for each scenario.
**Stick with Claude Code if:**
- It works for you, is paid for and available, and your projects are serious ones where reliability matters more than saving money.
- You are already used to the terminal and do not need a window.
- Your code contains sensitive data that cannot be sent to someone else's servers.
**Install and try ZCode if:**
- Claude is unavailable or expensive for you, and you need to work right now.
- You feel more comfortable in a window where you can see the files and the process than in a black terminal.
- You are just starting out and want generous free access to try vibe coding without spending anything.
My personal position is simple. ZCode is a good backup option and an excellent entry point for a beginner. I use Claude as my main tool, but keeping an inexpensive alternative at hand is sensible. Install it, run it for five days for free, and form your own opinion. It is worth trying.
## Why the tool itself is not the main thing
Key takeaways
Tools change every month: ZCode today, Grok yesterday, someone else tomorrow. Chasing the newest one leads nowhere. Consistent results depend on something else: how you put context into the agent's head—project architecture, a knowledge structure, a checking cycle. Set that up once, and any agent will work predictably.
I want you to take away a more important idea from this review than “GLM is a little weaker than Claude.” Over the past couple of months, a new Claude came out, Grok was updated, [open and free models](https://ailoot.ai/blog/en/free-coding-models/) caught up, and now there is ZCode too. Every month, someone declares themselves the best. If you chase every new release, you'll drown in switching and never build anything.
I have tested this on myself. My vibe coding failed most often for one reason: the model was not given context. The model itself has almost nothing to do with it; what matters is the setup around the agent. Give inexpensive GLM proper context, and it produces a working result. Without context, even the most expensive Claude produces a polished empty shell.
What stays the same as the tools change:
- **Project architecture.** How you divide a product into parts and what depends on what. You keep this in your head; the model has nothing to do with it.
- **The project's knowledge structure**—that second brain: files that give the agent context about your work instead of forcing it to guess.
- **The checking cycle.** How you review what the agent has done and catch mistakes before they make it into the product.
That is what context engineering comes down to: it does not matter which agent you have at hand; what matters is what context you give it. Set that up once, and ZCode, Claude, and any tool that comes along next will work predictably. Otherwise, you will keep changing interfaces every month, hoping the next one will finally “work by itself.”
### Sources
- [Z.ai - GLM Coding Plan](https://z.ai/subscribe)
- [Anthropic - Supported countries and regions](https://www.anthropic.com/supported-countries)
ZCode, Claude, this month's tool or next month's—it is all secondary. What comes first is the context you give the agent.
Get to know [Ailoot](https://ailoot.ai/).
## Questions about ZCode: a desktop alternative to Claude Code on GLM-5.2
### Does entering my own Claude key prove that ZCode sends no data to Z.ai?
No. The selected model endpoint, application proxying and telemetry are separate questions. A key identifies model access; verify the application’s documented routing and privacy behavior before using sensitive code.
### Are the GLM API prices charged per million words?
No. The quoted $1.40 input and $4.40 output rates are per million tokens. Compare actual input, output, cache usage and retries; a lower token rate does not establish a fixed saving per successful task.
## Ailoot · The company and its project portfolio
Ailoot combines established operating experience, a large client base, stable service and ongoing KPI control.
The company and its project portfolio: millions of customers; monthly sales in the billions.
Ailoot brings that operating experience to API access for AI features in customer products, agents, applications and automations.