ZCode is open source: Z.ai has published the complete ZCode workbench — desktop app, browser interface, backend services and terminal agent — in the zai-org/ZCode repository on GitHub under the Apache 2.0 licence, a release reported on 21 September 2026 as Z.ai opening ZCode's full coding agent stack to developers worldwide. So if you searched is ZCode open source hoping to inspect, self-host or build on a serious coding agent without a restrictive licence, the answer is yes — and the licence Z.ai chose is one of the most permissive available. The repository already shows 5.1 thousand stars and 1.4 thousand forks, so the community has clearly noticed.
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
Everything below comes from the zai-org/ZCode repository itself, checked on 21 September 2026, plus the same-day release coverage. Z.ai is the company behind the GLM model family, and ZCode is its AI programming workbench — the repository's own description calls it an AI programming workbench providing a desktop application, browser interface and terminal agent. This article covers what is actually in the repository, what the Apache 2.0 licence lets you do with it, and the one thing the release does not include.
Is ZCode Open Source? What the Repository Actually Contains
The release is not a stripped-down community edition — it is the full monorepo with four entry points, each of which you can run yourself:
- The desktop app: an Electron application, started in development with a single pnpm command, giving you the full graphical workbench on your own machine.
- The web and CLI workbench: a combined terminal-and-browser mode that bundles the text UI, web client, backend and agent runtime together — the closest thing to the hosted ZCode experience, self-hosted.
- The agent CLI: a terminal-based zcode command with the agent runtime that also powers the desktop and web versions.
- The server: HTTP and WebSocket services that manage connections between the clients and the agent runtime.
Structurally, the monorepo splits into a desktop package (Electron main, host and renderer processes), a web package for the browser client, a server package for the HTTP and WebSocket layer, and the zcode-cli app for the terminal agent. That layout matters if you plan to build on it: the agent runtime is not welded to any single interface, so you can take the piece you need and leave the rest.
What the Apache 2.0 Licence Lets You Do
Apache 2.0 is a permissive licence, and the practical consequences are worth spelling out. You can use ZCode commercially, modify it, redistribute your modified version, and run it for client work — all without paying Z.ai or open-sourcing your own changes. The licence also includes an explicit patent grant from contributors, which is precisely the protection that makes cautious companies comfortable adopting a tool. Your obligations are light: keep the licence and notice files intact and mark significant changes you make. Compared with the source-available licences some AI vendors ship — where commercial use or competing services are restricted — this is the clean end of the spectrum.
Licensing questions around AI coding tools rarely have answers this tidy. The community has been through this before with other agents: the is Jev open source question has a far messier answer, which makes a genuinely Apache-licensed full workbench stand out. If you are keeping score of which agent platforms you can actually own, ZCode just joined the short list.
If you want to go from cloning repositories to actually shipping agent-powered projects that pay, the AI Profit Boardroom has the setups, prompt libraries and weekly live coaching to get you there. Want a 1-on-1 plan for your specific business first? Book a free SEO strategy session and get a roadmap.
How to Run ZCode Yourself
The prerequisites are documented in the repository: Git, Node.js 24.14.0 and pnpm 10.33.2. From a clone, a pnpm bootstrap command sets up the workspace, after which you choose your surface — the desktop development command for the Electron app, the web development command for the combined browser-and-terminal workbench, or the CLI filter command for the terminal agent on its own. Note the Node version: this is a current-generation toolchain, and older Node installs will need upgrading before anything builds.
A sensible first session looks like this: clone the repository, run the bootstrap, start the web workbench, and give the agent one small, real task in a throwaway project — something with a test you can run, so you can watch how the agent handles feedback rather than judging it on a single generation. The terminal agent on its own is the lightest way in if the full workbench feels heavy, and because every surface shares the same runtime, whatever you learn about the agent's judgement in one mode carries over to the others.
Fair warning on expectations: running a development build of a workbench like this is closer to setting up a self-hosted platform than installing an app. If that sounds like more plumbing than you want, the takeaway of the release is still useful — the code being public means the hosted product you might pay for is inspectable, and the community can patch and extend it rather than filing feature requests into the void. The build your own OpenClaw guide covers the same self-hosting trade-offs for a different agent stack, and most of its advice transfers directly.
What This Means for the Coding Agent Landscape
Open-sourcing a complete agent workbench raises the bar for everyone. The reference points are familiar: Claude Code sits at the polished commercial end — the Claude Code free guide covers what you get there without paying — while community-driven stacks like OpenClaw have proven how fast open agent ecosystems can move, as the OpenClaw desktop app guide and the Hermes vs OpenClaw comparison both show. ZCode entering that landscape with an Apache licence and all four surfaces — desktop, web, terminal and server — gives builders a third serious option that is neither a closed product nor a loose collection of scripts.
What the release does not include is a model. The repository is the workbench and agent runtime; the intelligence you point it at is still a separate decision, whether that is a Z.ai model or another provider you configure. Treat the open-source announcement accordingly: you now own the tooling, not the brain. For how the different brains actually compare when run through identical hands-on tests, the Goldie Bench write-up covers that side of the decision, and it is the half of the stack where the real quality differences live.
The lock-in question is where an Apache-licensed workbench earns its keep. When your coding agent is a closed product, every workflow you build on it — custom prompts, project conventions, integrations — is a sunk cost you lose if pricing changes or the product pivots. With the workbench code public, the worst-case scenario softens considerably: the community can maintain a fork, your setup keeps working, and the vendor has to keep winning your subscription on merit rather than on switching costs. That does not make self-hosting the right choice for everyone today, but it changes the risk calculation for anyone betting serious workflow on the tool.
There is also a community angle worth naming. Tools with permissive licences accumulate plugins, forks and integrations at a pace closed tools cannot match, because every itch gets scratched by whoever has it. If you want to watch that dynamic — or contribute to it — the best OpenClaw community round-up shows what a healthy open agent community looks like in practice, and ZCode's early star count suggests it is about to grow one of its own.
The Verdict, and the Sensible Next Step
Yes, ZCode is open source — genuinely, fully, and under the licence you would hope for. The complete workbench is public, the Apache 2.0 terms allow commercial use and modification, and the only missing piece is the model access you were always going to need anyway. For tinkerers, it is a readable reference implementation of a modern agent workbench. For businesses, it is a coding agent platform you can adopt without licence risk. For everyone else, it is healthy pressure on every closed competitor to justify what the closed part is buying you.
If you are building towards a full agent operating setup, slot this alongside the Agent OS approach: the workbench is one layer, and the systems, prompts and workflows you run on top of it are where the compounding value comes from. Clone the repository, run the bootstrap, and give the terminal agent one real task — that single test will tell you more than any launch-day thread.
If you want the shortcut — working agent stacks, the prompts that drive them, and live help when you get stuck — check out the AI Profit Boardroom and skip the months of solo debugging. And for a personal plan on turning AI skills into revenue, book a free SEO strategy session — it costs nothing and saves you weeks.
