ZCode

A coding agent you run yourself: the same agent as a terminal CLI, a browser tab or a desktop app

Z.ai's coding agent harness. Powerful, intelligent, extensible.

No video published yet. The write-up below covers what the tool does and how to try it.

Category
AI agents & assistants
Audience
Developers
Language
TypeScript
Licence
Apache-2.0

Published Updated

ZCode is Z.ai's open-source coding agent workbench: one repository that ships the same agent as a terminal CLI, a browser app and a desktop client. It is for developers who already use a coding agent daily and want the harness itself — client, backend service, shared UI and agent runtime — in a codebase they can read, build and change. Teams choosing a harness rather than just using one are the natural readers.

What it does

Typing zcode with no arguments opens the terminal agent, which is the default. Adding a single flag brings the same agent up as a full browser app. Both run on your own machine, and the browser mode needs no Electron — the Electron desktop client is a separate front end living in the same repository.

Beyond the two front ends, the repository holds the backend service, a shared UI layer, and the agent CLI and runtime under apps/zcode-cli/. That is a plain directory, not a git submodule, so an ordinary clone gets everything.

There is also a workflow engine, and its notable trick is that you can raise a running task's concurrency limit without stopping the task. Release notes additionally say workflow scripts now burn fewer tokens; that is the project's own claim, not an independent measurement, and the material gives no number. The saved README records an update to ZCode v3.14.3 on 23 September 2026.

How it works

It is a TypeScript pnpm workspace monorepo, and every command runs from the repository root. pnpm bootstrap installs the workspace dependencies, prepares the desktop app's local runtime resources, then runs build:bootstrap. Narrower entry points exist when you do not want all of it:

  • pnpm install — dependencies only
  • pnpm prepare:desktop-runtime — desktop runtime resources, remote assets included by default
  • pnpm prepare:remote-assets — remote assets alone
  • pnpm bootstrap:with-remote — dependencies plus local and remote assets, serial build, skipping the desktop bundle
  • pnpm build — each workspace package's own build script

For development, pnpm dev:desktop prepares local resources, builds the desktop agent, then starts Electron with a source watcher. It defaults to production service configuration; pnpm dev:desktop:test points at the test environment. Setting ZCODE_DATA_BASE_DIR gives a dev build its own data directory. The remote SSH and WSL features need pnpm bootstrap:with-remote first to prepare a mock CDN, and when connecting to a remote project you choose the "download locally, then upload" asset option.

Getting started

The repository is github.com/zai-org/ZCode under Apache-2.0. It needs Git, Node.js 24.14.0 and pnpm 10.33.2, with mise.toml as the authoritative version source — those pins are exact, so a stock system Node probably will not do. Clone the repository, then run pnpm bootstrap.

The saved README covers development and packaging only; it points at no Releases page and no package-manager install, so treat this as build-from-source until the project says otherwise. The shell examples are macOS and Linux, and Windows appears only through the remote WSL support. No hardware requirements are listed anywhere in the material — no memory, VRAM or chip figures.

On money: the code is free under Apache-2.0, but a coding agent needs a model behind it, and the material does not say which model, whether a Z.ai account or API key is required, or what inference costs. The dev scripts do default to production service configuration, which implies a hosted service in the loop. Assume you pay someone for the intelligence and check that before committing a team. One practical note: the default README is in Simplified Chinese, with an English translation at README.en.md.

When to use it / when not

The catch that matters is a trust problem, not a rough edge: an earlier ZCode build silently uploaded Git history, and Z.ai apologised for it. If you work on code that cannot leave the building, pin a commit you have reviewed and watch what the agent sends. The repository is young as well — created on 20 September 2026 — so expect churn in the setup commands above.

Use it instead of Claude Code when you want one install that gives you both a terminal agent and a browser UI, and when you want the harness source in front of you. Don't use it if you need a signed installer and a vendor you can hold to a security promise today.

Alternatives

The honest comparison set is the other terminal coding agents and IDE-embedded assistants you already use; the material makes no benchmark claim against any of them. ZCode's distinguishing points are the two front ends from a single install and the mid-run concurrency change in its workflow engine. If you only want answers about code rather than an agent that acts on a repository, a plain chat window stays cheaper and simpler.

ZCode deserves a serious look from developers who want to see how a coding agent harness is actually assembled, and from anyone who would rather have a self-hosted browser UI than another Electron window. It is a harder sell in a locked-down shop: the Git-history incident is on the record, and the project does not yet document what it needs from a model provider or what that will cost. Clone it, read apps/zcode-cli/, try it against a repository you would not mind leaking, and judge from there.

More in AI agents & assistants

All of AI agents & assistants →