Z.ai, the Beijing-based AI lab formerly known as Zhipu AI, just launched ZCode, a free desktop application it calls an “Agentic Development Environment.” It’s purpose-built for their GLM-5.2 model, VentureBeat and it lands at a moment when the entire coding tools market is rethinking what “safe” means.
ZCode isn’t another terminal agent like Claude Code or Codex CLI. It’s a full desktop app for Windows, macOS, and Linux (Linux is still in beta) that puts the AI agent conversation at the center, with a file manager, terminal, Git panel, and live browser preview arranged around it. You describe what you want the agent to build, and it plans, codes, checks its own work, and keeps iterating until the objective passes verification.
And it’s free to download. The revenue comes from the GLM Coding Plan subscriptions, which start at $18 a month for Lite and go up to $160 for Max. There’s also a 5-day free trial with 5 million tokens per day with no credit card required.
What makes ZCode different from Cursor, Claude Code, and Copilot
I’ve used most of the major AI coding tools this year, and the first thing that stands out about ZCode is that it’s not trying to be VS Code with an AI sidebar. It’s an agent-first environment where the conversation with the AI is the primary interface, not an add-on.
[image: ZCode desktop app interface showing task management, code search, and AI agent progress tracking]
The core workflow is built around what Z.ai calls Goal Mode. You declare an objective with /goal <what you want done> and the agent works in rounds. At the end of each round, it runs automatic verification checks. If the objective isn’t met, it starts another round without waiting for manual approval.
You can steer mid-flight with commands like /goal replace and /goal pause, but the default loop is self-driving until the success criteria are satisfied.
The model underneath is GLM-5.2, a 753-billion-parameter mixture-of-experts model with about 40 billion active parameters per token and a 1-million-token context window. It’s the same model I covered when it launched last month, and the same one that beat Claude Code on security benchmarks just five days ago. Its API pricing runs about $1.40 per million input tokens and $4.40 per million output, roughly an 80 percent discount compared to Anthropic’s Opus 4.8.
GLM 5.2 launch coverage GLM 5.2 security benchmarks
The features that actually matter
ZCode has been shipping fast: eight public releases between June 26 and July 3, and it’s already at version 3.2.4. The feature set is surprisingly complete for a v1 product.
Custom subagents landed in v3.2.0 on June 29. You can define your own agents with specific models, tool permissions, and system prompts, stored as plain Markdown files at ~/.zcode/agents/. The primary agent auto-selects the right subagent based on task description, or you can invoke one explicitly with @agent-name.
The ability to pin a cheaper model to a read-only research agent while reserving GLM-5.2 quota for the code-writing agent is the kind of practical detail that suggests Z.ai has been dogfooding this internally.
The Remote Control feature is genuinely useful. Click a phone icon in the sidebar, scan a QR code, and your phone pairs to the currently open desktop workspace. You can check status, send instructions, approve file changes, and forward screenshots into the session. The desktop stays the runtime environment. The phone is a remote control, not a cloud instance.
The Bot Channel goes further, pairing ZCode with WeChat or Feishu so you can drive the agent conversationally from a messaging app. Discord, Slack, and Telegram support are listed as “coming in later versions.”
ZCode also supports bring-your-own-key for six providers beyond Z.ai: Anthropic, OpenAI, OpenRouter, Moonshot (Kimi), MiniMax, and any OpenAI/Anthropic-compatible endpoint. And it can one-click import MCP configs from Claude Code, Codex CLI, and OpenCode, which is a nice touch for anyone who’s already invested in the MCP ecosystem.
The pricing story: free app, subscription model, and the July 31 cliff
The app itself is free. First-time users get a 5-day trial with 5 million tokens a day, no credit card required. After the trial, agent features require either a GLM Coding Plan subscription or a pay-as-you-go API key.
The Coding Plan tiers are Lite at $18 a month, Pro at $72 (5x the usage), and Max at $160 (20x the usage). Current pricing reflects about a 10 percent introductory discount, so expect those numbers to edge up after the promo period.
There’s also a ZCode-specific quota promotion running through July 31. During off-peak hours (20 of 24 hours daily), token consumption is metered at 0.67x instead of the standard 1x. During peak hours (2 PM to 6 PM daily), it’s metered at 2x instead of 3x. The net effect is roughly 1.5x effective quota while the promo lasts.
On August 1, the coefficients revert. If you’re evaluating ZCode with the intention of subscribing, budget against the post-promo numbers.
The GLM Coding Plan isn’t limited to ZCode either. It works across 20+ coding tools including Claude Code, Cline, Kilo Code, and OpenCode, which makes it a viable alternative to paying Anthropic or OpenAI directly if GLM-5.2’s capabilities fit your workflow.
Why the Fable 5 export ban makes this release more important
ZCode launched about three weeks after the US government issued an export control directive that banned foreign nationals from accessing Anthropic’s Fable 5 and Mythos 5 models. The directive was lifted on June 30, but the damage to enterprise trust was done.
The episode introduced a new risk category into AI procurement: sovereign access risk. When a government can disable a commercially deployed AI model overnight, traditional evaluation criteria: developer experience, benchmark scores, pricing become secondary to a more fundamental question.
Z.ai’s response was strategic. On the same day the export controls were announced, Zhipu published GLM-5.2’s open-source weights under an MIT license with no usage restrictions. A development team can download the model, host it on its own infrastructure, and run ZCode against it without ever touching Z.ai’s cloud. That eliminates both US export-control risk and Chinese data-sovereignty concerns in a single move, provided you self-host.
This is the part that competing tools can’t easily replicate. Cursor doesn’t give you the underlying model to run yourself. Claude Code certainly doesn’t. GitHub Copilot is tied to Microsoft’s cloud. ZCode is the first agentic development environment where the entire stack (IDE, model weights, API) can be decoupled and self-hosted. That’s not going to matter to every developer, but for enterprises with compliance requirements, it’s a real differentiator.
Where ZCode still has room to grow
I haven’t spent time with ZCode myself yet, so I’m reporting what early users and reviewers are finding.
Linux support is still in beta. Some developers on X have reported that the Linux ARM64 build has known stability issues.
The Bot Channel currently supports only WeChat and Feishu. Discord, Slack, and Telegram are promised but not yet delivered, so if your team lives in Western messaging apps, this feature isn’t ready for you.
Custom subagents launched with a beta label on June 29, and while the feature is functional, it’s user-level only. There’s no per-project subagent configuration yet.
And security reviewers have flagged the need for careful evaluation of credential handling, particularly for the remote development and messaging-platform-triggered task paths. An agent that can be summoned from WeChat involves access patterns that should be mapped before trusting it with sensitive code.
The bottom line
ZCode is the most consequential coding tool launch since Cursor hit its stride because it represents something the market hasn’t had before: a vertically integrated stack from a model lab that can undercut Western pricing by an order of magnitude while offering an MIT-licensed fallback that eliminates regulatory kill-switch risk.
The app itself looks genuinely well-designed based on the documentation and early user reports. The near-daily release cadence suggests a team that’s shipping fast and responsive to feedback. And the underlying GLM-5.2 model is legitimately competitive: second on Code Arena behind only Fable 5, beating Claude Code on security benchmarks.
Whether Z.ai can build trust with Western enterprise buyers amid escalating tech tensions is an open question. But the company’s market cap crossing $128 billion suggests investors are betting yes. For individual developers, the risk is near zero: it’s free, it works with your own API keys, and the model weights are open. Download it, run the 5-day trial, and see if GLM-5.2 fits how you code.
