ZCode is Z.ai’s agentic coding IDE, launched July 2, 2026. It’s the desktop counterpart to GLM-5.2, the open-weight coding model that has been quietly undercutting Claude and GPT on price all year, and it’s the clearest sign yet that the cheap-strong-model trend now has a full product wrapped around it.

New in September 2026: ZCode uploaded users’ entire repositories, then open-sourced itself

On September 18, a researcher going by Ferstar documented that ZCode was silently packaging up whole user workspaces — full git histories, file caches, the lot — encrypting them and shipping them to an Alibaba Cloud bucket. Reports put one observed archive at 313MB across 564 upload attempts. The decryption key lived only on Z.ai’s server, so users could neither read what had been taken nor delete it. There was no setting to turn it off and no mention of it in the privacy policy. Z.ai traced it to the Codebase Indexing feature, which was on by default.

The response was faster and more thorough than the industry norm. Z.ai apologized, said the data was never used for training, removed the Repo Wiki feature in client v3.14.0, and commissioned CAICT and NSFOCUS to audit the fix — both reported the bucket and its contents deleted. On September 21 the company open-sourced the entire ZCode client on GitHub and committed to a vulnerability reporting process with paid rewards. Ferstar confirmed the published code no longer contains the upload path, but criticized Z.ai for wiping the commit history, which means the pre-patch code is no longer inspectable.

This is the second time in three months a major vendor’s coding tool has been caught doing this — xAI’s Grok Build was found sending entire repos to the cloud in July. Read it as a category problem, not a China problem: agentic tools need to index your codebase to be useful, “index” and “upload” are one config flag apart, and defaults are set by whoever ships fastest.

What it is

A free, cross-platform (Windows/macOS/Linux) desktop app in the same shape as Cursor or Claude Code: an agent plans, edits across files, runs commands, and iterates. The differences are pricing and openness.

Price. The free tier is genuinely usable. Paid Lite runs $16.20/mo and Max $144/mo, against roughly $20–$200 for the Western equivalents, and there’s a 30% promo through September that drops Lite to about $12.60. If you already pay for a GLM Coding Plan, that quota carries into ZCode.

BYOK. You can point ZCode at your own API keys and run Claude or GPT through it instead of GLM-5.2. That matters more than it sounds. It means adopting ZCode isn’t a bet on Z.ai’s model staying good, just on the harness being decent.

Open weights. GLM-5.2 is MIT-licensed, so an enterprise can self-host the model and keep everything on its own infrastructure. Very few tools in this category offer that.

There’s also a genuinely novel feature: you can trigger and steer coding tasks remotely over Telegram, WeChat, or Feishu. Kick off a task from your phone, get pinged when the agent needs a decision. It’s the kind of thing that sounds gimmicky until you’ve spent a Sunday waiting on a long agent run.

The thing you need to weigh

If you use ZCode’s hosted API rather than BYOK or self-hosting, your code and prompts are processed under Chinese data law. That isn’t a slur on the product. It’s a fact about jurisdiction, and it’s the first question your enterprise customer, your insurer, or your acquirer will ask. For a solo founder building a side project, it’s a non-issue. For anything touching health data, EU personal data, US government work, or a codebase you plan to sell, it’s disqualifying unless you self-host or BYOK. Decide which of those you are before you get attached to the price.

The non-coder reality

ZCode is an IDE. It assumes you can read a diff, run a terminal command, and tell when the agent has gone off the rails. Like Cursor and Claude Code, the agent makes you faster if you already know what “right” looks like; it makes you confidently wrong if you don’t. A rating of 2 is honest. This is not a tool for someone building their first app.

Bottom line

ZCode is still the best value in agentic coding on price, and it’s not close. The harness is credible rather than exceptional (Cursor still has the better editor, Claude Code the better agent judgment) but at free-to-$16 with BYOK and open weights, the value math is hard to argue with for anyone technical enough to use it.

The September incident cuts both ways, and we’re not going to pretend otherwise. Shipping a default-on feature that uploads your entire git history under a key you don’t hold is a serious failure of judgment, and wiping the commit history during remediation makes the cleanup harder to verify than it needed to be. Against that: Z.ai fixed it in three days, paid for two outside audits, published the client source, and opened a disclosure process. Most vendors in this category have done less after worse.

The decision point hasn’t changed, it’s just sharper. If your code can live under Chinese data law and you’re comfortable reading a config before you trust a default, ZCode is a strong buy. If it can’t, or you aren’t, the price was never the point. Sources: VentureBeat, Tech Times, ChatForest builder guide.

Was this helpful?