The harness, not just the model
Z.ai, the company behind the GLM model family, open-sourced ZCode this week, and the interesting part is the scope. This is not a CLI script and a README. The repository contains the desktop client, the browser workspace, the backend services, the shared UI, and the agent runtime that powers all of it, under the Apache 2.0 license. Three days after the repo went public it had crossed 6,300 stars, which tells you how starved this space is for a first-party harness people can actually read.
Most model companies ship an API and let other people build the tooling around it. Z.ai went the other way: they built the harness first, tuned it around their own model, and then published the whole thing. The repo description calls it their coding agent harness, powerful, intelligent, extensible, which is marketing, but the code is real and checkable, which matters more.
What is actually in the repo
ZCode ships three interfaces from one codebase. There is an Electron desktop application for people who want a workspace, a web and terminal mode that packages the TUI, web client, backend, and agent together, and the zcode agent CLI, which is also the runtime that the desktop and web interfaces sit on. The agent CLI lives in apps/zcode-cli inside the repo, no submodules required. Development is TypeScript heavy, close to 33 million lines of it according to the GitHub language breakdown, with some C# and Swift for platform packaging.
The setup path is ordinary for a pnpm monorepo: Node 24, pnpm 10, a mise.toml file pinning tool versions, and a pnpm bootstrap command that installs dependencies and builds the packages. Anyone who has built a VS Code fork or an Electron app before will find the shape familiar. That ordinariness is a feature. You can clone it, read the agent loop, and see exactly how the tool executes commands and asks for confirmation before touching anything sensitive.
Why a model company publishes its harness
The official docs frame ZCode as an agentic development environment built specifically for GLM-5.3, and that framing is the real story. The docs claim GLM-5.3 runs with a stable 1 million token context, and ZCode is tuned to exploit it: the agent keeps goals, file state, terminal output, and git status in one long-horizon task instead of restarting every few steps. The docs also describe confirmation gates for sensitive commands and file changes, and remote follow-up channels through desktop, mobile, and Feishu or WeChat bots, so a long task keeps running while you poke at it from your phone.
The competitive logic is worth naming. Claude Code, Codex CLI, and Gemini CLI are all closed-source harnesses tied to their vendor’s models. OpenAI and Anthropic treat the harness as a moat. Z.ai is betting the opposite way: if the harness is open, third parties can integrate it, competitors can fork it, and the value concentrates in the model instead. Apache 2.0 makes that explicit. Whether the bet works depends on whether GLM-5.3 holds up in independent benchmarks, but the strategic move is easy to read, and users benefit either way.
What this means for developers
For anyone running coding agents daily, two things follow. First, there is now a credible open harness you can self-host and modify, which matters if you want custom tool integrations, on-prem execution, or audit trails your compliance people can actually inspect. Second, harness behavior is model-specific in ways vendors rarely admit. ZCode is explicitly tuned for GLM-5.3, and the docs say so. Expect the same coupling elsewhere: the harness and model are one system, and benchmark numbers quoted for one do not transfer to the other running under a different harness.
It also lowers the cost of experimentation with agentic workflows. If you want to see how a production coding agent structures planning, permission checks, and review, the source is right there. That was previously true only for community projects like Aider or OpenHands. Now a major model vendor is in that set.
Trying it without a big commitment
The cheap path: clone the repo, run pnpm bootstrap, and start the terminal agent in TUI mode against a small project before pointing it at anything real. You will need access to GLM models, either through Z.ai’s coding plans or your own API keys, so price that before you get attached. If the terminal experience holds up, the desktop app and web UI are the same runtime underneath, so nothing gets thrown away when you switch interfaces.
My take: I would not read this as Z.ai beating anyone. It is a young repo with one push day so far, and harness quality is only proven by months of daily use. But the move itself is notable. Open-sourcing the entire agent stack puts pressure on every vendor that charges for a closed harness, and it gives developers something concrete to read instead of marketing pages. The repo to watch is the agent runtime in apps/zcode-cli, because that is where the actual agent loop lives, and it is the part everyone else currently keeps to themselves.