Zhipu publishes ZCode source following a data-handling dispute

What changed after the data dispute

Zhipu has made ZCode's source code public and apologized following concerns about repository data leaving users' machines. IT Home's September 21 account says the issue involved Repo Wiki, a feature whose page-generation process could trigger repository uploads and had initially been enabled by default.

According to the company's statement reproduced by IT Home, version 3.14.0 removed the Repo Wiki entry point and the local snapshot-generation and upload path. Zhipu said the relevant cloud storage had been cleared and reported checks by the China Academy of Information and Communications Technology and NSFOCUS. These are reported remediation statements; China AI Wire has not independently audited the deployed application.

What the repository makes inspectable

The ZCode repository describes a workbench spanning a desktop application, a browser interface and a terminal agent. It contains the client, backend, shared interface code and agent runtime. Its README now records a September 23 update to version 3.14.3.

Developers can inspect how the local application connects to services and can build components themselves. The documentation also distinguishes local web use from network exposure: the web mode listens on the loopback address by default, while a non-local host normally receives an access token. These details matter because the security of a local development tool changes when its interface becomes reachable from another machine.

What is open, and what still needs configuration

The repository carries an Apache-2.0 license and documents builds for desktop and command-line use. Its web and terminal modes can run without the Electron desktop shell. It also separates shared interface packages, backend services, provider integrations and the agent runtime, giving reviewers more than a screenshot or a high-level architecture diagram to inspect.

The README says the standard desktop development command uses production-service configuration unless a test mode is selected. That is a concrete reason to review configuration as well as source: running a local build does not necessarily mean every connected service is local.

Open application code should also be distinguished from open model weights. The repository exposes the harness and related software; a selected model provider may still operate remotely under its own terms. A local interface can send a request to an external model, so location of the user interface alone cannot establish where all data is processed.

The project notice spells out remaining access

ZCode’s published project notice states that its shared agent execution adapter does not provide an operating-system sandbox by default. Tools can work within the permissions of the operating-system account that runs them. In remote workspaces, actions may occur on an SSH host, in WSL or inside a container, rather than on the computer displaying the interface.

The same notice distinguishes removal of a repository-upload feature from the network requests needed to operate an AI agent. Model calls can include prompts, conversation history, code, tool results or attachments, depending on the function used. It also describes gateway forwarding for specified official Coding Plan endpoints. These are disclosures in the project’s documentation; this article has not performed a traffic capture or audited its servers.

That makes the immediate configuration question specific: which model endpoint receives a request, which machine executes the tools, and what account permissions that machine grants. Reviewing those three settings is necessary even when a developer builds the newly published source locally. The September remediation concerns a particular upload path; it does not convert a connected coding agent into an offline program.

Source availability and data policy are separate

Publishing code improves visibility into a particular codebase. It does not prove that every installed binary matches that revision, that every server has identical settings, or that all historic data was deleted. A review needs to connect the release version, network behavior and service policy.