ZCode is not another terminal coding agent. It is a full desktop application. You do not type commands into a terminal prompt. You open an app, define a Goal, and watch an AI agent plan, execute, and verify multi-step coding tasks across your project.
Built by Z.ai (Zhipu AI), the same company behind the GLM model family, ZCode launched in June 2026. It runs on Mac, Windows, and Linux (beta). It is powered by GLM-5.2, a 744-billion parameter mixture-of-experts model with a 1-million token context window. And it has features no other coding agent offers: remote control from your phone, SSH remote development, and a structured Goal system for complex objectives.
This is everything you need to know.
Who Makes ZCode?
Z.ai is the product arm of Zhipu AI, the Chinese AI company behind the GLM model series. This matters because ZCode is a first-party tool from the model maker. The team that built GLM-5.2 also built the agent that uses it. This is the same relationship Anthropic has with Claude Code, or that OpenAI would have if they shipped a coding agent.
First-party integration means the agent is optimized specifically for its model. Prompt engineering, tool use patterns, and context management are all tuned for GLM-5.2’s strengths and weaknesses.
For more on the GLM model itself, see our GLM-5.2 complete guide. For the broader Chinese AI model landscape, check best Chinese AI models 2026.
Desktop App, Not Terminal
This is the most important distinction. Claude Code, MiMo Code, Aider, and OpenCode all run in your terminal. ZCode is a native desktop application with a graphical interface.
What does this mean in practice?
- Visual project navigation (file trees, search panels)
- Rich diff views for code changes
- Checkpoint visualization for multi-step tasks
- Integrated terminal within the app
- Point-and-click configuration instead of config files
- Drag-and-drop file selection
If you prefer a visual interface over terminal commands, ZCode feels very different from the terminal-native agents. If you love your terminal workflow, ZCode will feel heavy and opinionated. Neither is wrong; they serve different developer preferences.
For a direct comparison of the GUI vs CLI approaches, read ZCode vs Claude Code.
The Goal System
ZCode’s most distinctive feature is the Goal system. Instead of giving the agent one command at a time, you define a Goal: a complex objective that may take many steps to complete.
Examples of Goals:
- “Add user authentication to this Express app using JWT with refresh tokens”
- “Migrate the database layer from MySQL to PostgreSQL”
- “Implement a complete CRUD API for the products resource with validation, tests, and documentation”
When you submit a Goal, ZCode:
- Plans: Breaks the Goal into discrete steps
- Shows you the plan: You review and approve before execution begins
- Executes: Works through each step, reading files, writing code, running commands
- Checkpoints: Saves state at each step so you can review, revert, or branch
- Verifies: Runs tests or checks at the end to confirm the Goal is met
This is fundamentally different from the conversational back-and-forth of terminal agents. You do not need to guide the agent through each step. You state the objective, approve the plan, and let it work.
The Goal system excels for well-defined, multi-step tasks. It is less ideal for exploratory work where you are figuring out what you want as you go. For exploration, ZCode still supports a conversational mode, but that is not its strength.
Remote Control
This one is genuinely unique. No other coding agent offers anything like it.
ZCode can be controlled remotely via:
- WeChat (popular in China)
- Feishu (Lark, used by many companies)
- Telegram (global)
You connect your messaging app to your ZCode instance. Then, from your phone, you can:
- Start new Goals
- Check progress on running Goals
- Approve or reject agent decisions
- Receive notifications when tasks complete or need input
- Steer the agent with additional instructions mid-task
Use cases:
- Start a long build task before you sit down at your computer
- Monitor progress during lunch
- Approve a deploy step from your phone
- Receive “task complete” notifications without watching the screen
We wrote a dedicated tutorial on ZCode remote control via Telegram with full setup instructions.
SSH Remote Development
ZCode can connect to remote servers via SSH and work against remote codebases. This is not unique (VS Code does this too), but in a coding agent context it opens interesting workflows:
- Develop against a powerful remote machine from a lightweight laptop
- Work directly on staging servers for debugging
- Access codebases that cannot be copied locally (security requirements)
- Use remote GPU instances for ML projects
You configure SSH connections in ZCode’s settings. Once connected, the agent reads and writes files on the remote machine, runs commands remotely, and manages everything as if the code were local.
Multi-Agent Coordination
Newer versions of ZCode support multi-agent coordination. Multiple agent instances can work on different parts of a Goal simultaneously.
For example, if your Goal is “add authentication and write tests,” ZCode might spawn:
- Agent 1: Implements the auth module
- Agent 2: Writes tests for the auth module in parallel
The agents are aware of each other and coordinate to avoid conflicts. This is still maturing, but for large Goals that decompose into independent sub-tasks, it meaningfully reduces completion time.
Integrated Tools
ZCode bundles 20+ programming tools directly into its interface:
- Git (commit, branch, merge, history)
- Terminal (command execution)
- File system operations
- Package managers (npm, pip, cargo, etc.)
- Build tools
- Test runners
- Linters and formatters
- Database clients
- API testing
You do not need to switch between ZCode and external tools for most development tasks. Everything lives within the app.
In-Workflow Code Review
At each checkpoint during Goal execution, ZCode pauses and shows you what changed. You review the diff, approve or reject, and the agent continues or adjusts.
This is more structured than terminal agents where you review changes after the fact (or miss them entirely). The checkpoint system means you stay informed and in control throughout multi-step operations.
Pricing
ZCode requires a GLM Coding Plan subscription from Z.ai. Pricing ranges from approximately $10 to $30 per month depending on usage tier.
This is middle ground pricing. Cheaper than Claude Code’s Max plan for heavy users. More expensive than MiMo Code (which is free). You get the full desktop app, GLM-5.2 model access, remote control features, and multi-agent coordination.
There is no free tier announced. If cost is your primary concern, MiMo Code offers free access to a frontier model. If you value ZCode’s unique features (Goals, remote control, desktop GUI), the subscription is reasonable.
How It Fits the GLM Ecosystem
ZCode is the coding-specific application of GLM-5.2. Z.ai positions it as:
- GLM-5.2: The foundation model (MIT-licensed weights, open for any use)
- ZCode: The developer-facing product built on GLM-5.2
- GLM Coding Plan: The subscription that funds development
If you use GLM-5.2 with other tools (like Claude Code with GLM-5.2 or MiMo Code with GLM), you get the raw model. ZCode gives you the agent layer on top: goal planning, execution, checkpoints, remote control.
GLM-5.2 itself is a 744B MoE model with impressive benchmarks. The fact that its weights are MIT-licensed means anyone can build on it. But ZCode represents Z.ai’s own optimized implementation.
Strengths
- Goal system is genuinely powerful for complex, multi-step tasks
- Remote control is unique in the market and useful for mobile-first workflows
- Desktop GUI is more accessible than terminal tools for many developers
- First-party model integration ensures optimized performance
- SSH remote development expands where you can work
- Checkpoint system keeps you in control during long operations
Weaknesses
- Cost: Not free, requires monthly subscription
- Desktop only: No mobile version, no web version (remote control is the phone story)
- GLM-5.2 only: Cannot swap model backends (unlike MiMo Code)
- Linux support in beta: Not fully stable on all distros
- Newer ecosystem: Fewer tutorials, community resources, and integrations than Claude Code
- Less flexible for exploration: Goal system is great for defined tasks, less natural for open-ended tinkering
Who Should Use ZCode?
ZCode fits developers who:
- Prefer graphical interfaces over terminals
- Work on complex, multi-step features regularly
- Want remote control to manage tasks from their phone
- Develop against remote servers via SSH
- Are already in the Zhipu/GLM ecosystem
- Value structured execution with checkpoints over freeform conversation
If you live in the terminal and prefer conversational coding, Claude Code or MiMo Code are better fits. If you want a visual, structured, goal-oriented workflow, ZCode offers something nobody else does.
For the full comparison, see ZCode vs Claude Code or our three-way comparison.
Getting Started
Ready to try it? Our ZCode setup guide walks through installation, configuring your GLM Coding Plan, creating your first Goal, and setting up SSH and remote control.
FAQ
Can I use ZCode without a subscription?
No. ZCode requires a GLM Coding Plan subscription (~$10 to $30/month). There is no free tier. The model access and agent features are bundled into the subscription.
Is ZCode open source?
No. ZCode is a proprietary desktop application. However, the underlying GLM-5.2 model has MIT-licensed weights, so you can use the model itself in other contexts. ZCode as the agent/app layer is closed source.
Can I use ZCode with models other than GLM-5.2?
Not currently. ZCode is designed specifically for GLM-5.2 and does not support alternative model backends. This is a trade-off: tight integration but no flexibility.
How does ZCode compare to Cursor?
Both are GUI-based coding tools, but they are different categories. Cursor is an IDE (code editor with AI features). ZCode is a standalone agent (defines and executes Goals independently). Cursor helps you write code line by line. ZCode completes entire objectives autonomously. See Claude Code vs Cursor for more on the IDE vs agent distinction.
Is the remote control feature secure?
Remote control works through authenticated connections between your messaging app and your ZCode instance. You authenticate the connection once, and subsequent messages are verified. However, anyone with access to your connected messaging account could theoretically control ZCode. Use secure messaging accounts and enable two-factor authentication on your Telegram/WeChat/Feishu accounts.