
ZCode review: Z.ai's GLM-5.3 coding agent on a real project
I tried ZCode with GLM-5.3 on a real project. Here's how the coding agent works, what it costs, its permission modes, automations, API, and limits.
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
I tried ZCode with GLM-5.3 on a real project. Here's how the coding agent works, what it costs, its permission modes, automations, API, and limits.
I installed it on my Mac and pointed it at this site’s repository. I wanted to see how it handled a real project, not a small demo.
The first result was good. GLM-5.3 found useful improvements in a codebase with more than 1,700 posts, and the app felt fast with a clear workflow.
ZCode is available for macOS, Windows, and Linux.
Z.ai calls it an Agentic Development Environment, or ADE. Unlike a VS Code fork, it has no traditional code editor.
You give the agent a task. It plans the work, edits files, runs commands, uses the browser, and checks the result.
task -> plan -> edit files, run commands, use the browser -> verify the result -> review
The task keeps everything together: your goal, changed files, terminal output, browser context, and Git state.
You can also run tasks in parallel, schedule recurring work, and control ZCode from your phone.
Z.ai is the international name of Zhipu AI. The Beijing AI lab grew out of research at Tsinghua University.
It builds the GLM family of models. Unlike OpenAI and Anthropic, Z.ai also publishes model weights you can download and host.
ZCode is its official harness for GLM-5.3. The app is built around the model, and the model is tuned for this kind of agentic work.
Z.ai sells API access and the GLM Coding Plan. The same subscription works in more than 20 coding agents, including Claude Code.
It uses the same base model as GLM-5.2. Z.ai did not run another pre-training phase. The improvements came from post-training on longer engineering tasks.
Reasoning is always active. You can reduce it, but you cannot disable it. Higher effort also uses more of your plan quota.
Z.ai trained GLM-5.3 inside task environments.
The model receives a goal, some tools, and an outcome it can verify. It writes code, runs it, and learns from whether the task passed.
Some environments represent several days of engineering work. A task might ask the model to find a bottleneck, fix it, run an experiment, and prove the improvement.
Creating enough realistic environments by hand would take too long. Z.ai uses research agents to collect work patterns and turn them into runnable tasks. Another agent checks that each task can be solved.
The verifiers do not see the reference solution. This makes it harder for the model to pass by copying the expected shape.
This training explains why GLM-5.3 works well inside ZCode. It is built to keep going through long tasks without asking you to split every step.
Z.ai reports a 50% improvement over GLM-5.2 on its internal coding benchmark.
I do not put much weight on one company benchmark. The public results show a similar improvement:
Z.ai says GLM-5.3 leads open models on Terminal Bench 3.0 and Agents’ Last Exam.
The comparison with closed models is less dramatic.
On Z.ai’s Code Bench, GLM-5.3 scores 31.4% at high effort. It produces about 50,000 output tokens per task. Claude Opus 4.8 scores 29.5% with about 120,000 tokens.
GLM-5.3 reaches 34.5% at max effort. Claude Fable 5 stays ahead at 39.5%.
The model does not lead the available coding models, but it gets close while using fewer tokens and costing much less.
GLM-5.3 became unusually good at finding security problems during training.
It scores 84.5% on CyberGym. Z.ai says this is the best result on that benchmark so far. Its exploitation scores also more than doubled over GLM-5.2, although closed models still lead.
Z.ai then worked with security teams on real projects. The model found 2,436 vulnerabilities across 269 projects. That included 1,097 medium-to-high severity issues.
Some bugs had been in those codebases for decades. Z.ai tracks the reports in a public Security Disclosure Ledger.
This capability delayed the public weights. GLM-5.3 went through another two weeks of risk review before release.
You can download the GLM-5.3 weights from Hugging Face. Z.ai published them on August 28, 2026.
GLM-5.2 used the MIT license. GLM-5.3 uses a custom license.
You can still use, modify, distribute, sell, deploy, and fine-tune the model. The extra condition applies to model-hosting companies earning more than $10 billion per year. They need a Z.ai security review.
The extra condition does not affect most developers, but open weights no longer means MIT licensed. Read the license before building a product around the model.
Z.ai released GLM-5.3-Flash two days before the main weights. Flash uses the MIT license.
It is the first natively multimodal GLM-5 model. It can inspect images and screenshots.
Flash has 320 billion total parameters, with 18 billion active for each token. Its mixture-of-experts design keeps inference cheaper.
It also combines sparse and linear attention. This lowers the cost of processing long contexts.
Flash costs about one tenth as much as GLM-5.3. Inside the Coding Plan, it gives you three times the usable quota.
ZCode uses it for visual work. The agent can open a page, inspect the result, and check its own changes.
ZCode sends your prompts and relevant code to Z.ai when you use the Coding Plan.
Z.ai is a Chinese company. Your company may have rules about sending proprietary code or client data to its infrastructure.
For public projects and personal experiments, I am comfortable trying it. I would check the policy before using it for private client work.
You can avoid the hosted service by running the open weights yourself. In practice, that needs serious hardware. We will come back to this later.
On an Apple Silicon Mac, you get a .dmg file. Open it and drag ZCode into Applications.
Windows installers are available. Linux builds are currently in beta.
A new AI Workshop cohort for $99, reduced from $299
My Black Friday offers open November 23. Join the new AI Workshop cohort for $99 instead of $299, or get all four TypeScript, React, Git, and AHA Stack masterclasses for $49. Join the separate email lists to hear when each offer opens.
ZCode asks you to connect an account on first launch.
For most people outside China, choose Z.ai.
New users currently get a five-day trial. It includes daily quotas for GLM-5.3 and GLM-5-Turbo.
After the trial, you need a GLM Coding Plan:
All three plans were 30% off when I checked them.
Usage works through points. Calls during off-peak hours and weekends cost half as many points.
The sidebar contains your projects and tasks. It also gives you access to search, automations, and the Plugin Marketplace.
The center of the app is one prompt box. Use @ to add context and / to run commands.
ZCode also includes a terminal. Click the terminal icon in the top-right corner to open it.
The model selector sits below the prompt box.
GLM-5.3 also asks you to choose a reasoning effort.
I use Max for architecture, debugging, and work across several files. Low is enough for renames, small edits, and questions.
If you are unsure, start low. You can run the task again with more effort. Spending a large part of your quota on a small edit is harder to undo.
I would check the permission mode before starting a task.
Start with Ask before changes. You will approve more actions, but you will see what the agent wants to do.
When you trust it on that project, switch to Edit automatically.
Use Plan mode for large changes. Reading a plan takes a minute. Finding the wrong assumption inside a large diff takes much longer.
I would keep Full access for disposable projects. Always use Git, regardless of the mode. My free Git course covers the basics if you need them.
I selected the flaviocopes.com repository and gave ZCode this task:
see the last published post and analyze if we can improve it, suggest up to 4 things I could do to make it better
The agent started reading the repository. I was using Ask before changes, so it stopped before running a shell command.
I could allow the command once, allow that pattern for the project, or deny it. This lets you grant access gradually instead of approving everything upfront.
I approved the command. ZCode found the latest post, read it, and compared it with the rest of the site.
It found dead-end links, a long section without a table of contents, and a timezone inconsistency.
That was useful work from the first task.
The prompt also helped. I gave it one goal and a clear limit. With coding agents, I try to:
A normal task is a conversation. A Goal keeps working toward an outcome.
ZCode loops through planning, execution, and verification. It updates its plan while it works instead of stopping at the first answer, which is where GLM-5.3’s long-task training should help.
I would still give Goals an outcome you can check. “Make the test suite pass” works. “Make the app better” does not.
The Plugin Marketplace adds new capabilities.
Plugins can bundle skills, commands, and MCP servers. The current catalog includes browser automation, Android emulation, video tools, and document production for DOCX and PDF.
If MCP is new to you, read What is MCP, the Model Context Protocol.
You can schedule recurring tasks or queue work for later.
The included templates cover a morning development brief and a daily risk scan. There is also a release brief and a documentation sync check.
I would use automations for work that does not need an immediate result:
The security scan is the first one I would try. Vulnerability discovery is one of GLM-5.3’s strongest areas.
Remote mirrors the desktop app on your phone after you scan a code.
Bot Channel connects ZCode to WeChat, Feishu, or Telegram. You can check a task, start another one, change projects, switch models, and select a run mode from chat.
The session still runs on your desktop. Your code stays on that machine.
I like this idea because long agent tasks often stop to ask one question. I would rather answer from my phone than find the task blocked two hours later. For Codex, Cursor, Grok Bot, and SSH workflows, see How to work with coding agents from your phone.
Yes. On September 20, 2026, Z.ai published ZCode’s source code on GitHub under the Apache 2.0 license, and the repository includes a command line version.
It’s one zcode command. With no arguments it opens a terminal UI. With --web it serves the same agent as a local web app:
zcode zcode --web
Both run on your machine without the desktop app. The ZCode website only offers the desktop download, though, so you build the CLI from the repository with Node.js and pnpm, following the README.
There’s also zcode-app-cli, a community package that gives you a zcode command with npm install -g zcode-app-cli@latest. Z.ai doesn’t make it. It wraps the agent from a ZCode desktop release in its own terminal interface.
Be careful with one default in the official CLI. When you run a task without the interface, with --prompt, it uses the yolo permission mode unless you pass --mode, so normal tool calls run without asking.
If you want a terminal agent on the GLM Coding Plan without building anything, Claude Code is the simplest route.
You are not locked into ZCode. Z.ai documents the Coding Plan for more than 20 tools.
The fastest way to configure Claude Code is Z.ai’s helper:
npx @z_ai/coding-helper
You can also edit ~/.claude/settings.json yourself:
{ "env": { "ANTHROPIC_AUTH_TOKEN": "your-coding-plan-api-key", "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic", "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3", "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.3", "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.3-flash", "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000", "API_TIMEOUT_MS": "3000000" } }
The three model settings route Claude Code to GLM. The compact window allows it to use more of the 1M context before summarizing.
The long timeout prevents a high-effort request from ending too soon.
Use the exact endpoint shown above. A wrong base URL can produce 1113 Insufficient Balance, even when the subsc
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。