Z.ai's new Agentic Development Environment (ADE) skips the "assistant in a sidebar" model entirely. You hand it a goal; it drives the file edits, terminal, git, and browser preview itself. This brief covers what it is, what it costs, and how it actually stacks up against the six tools it's being measured against.
The core pitch is a shift in who drives. A traditional IDE waits for keystrokes; ZCode's agent takes a natural-language goal and keeps working across a whole task thread.
Z.ai frames this as long-horizon execution: the agent recovers progress across a session rather than starting cold each time. It's a desktop app for macOS, Windows, and a Linux beta, and it's marketed explicitly as the "official" way to bring the GLM-5.2 model family into a coding workflow — not a general-purpose editor that happens to support one more model.
GLM-5.2 is the default and the reason ZCode exists — but the door to other models is deliberately left open.
The agent is described as purpose-tuned for GLM-5.2's long-context and long-horizon behavior. GLM-5-Turbo is available as a lighter option depending on account tier.
Provider settings accept a custom API key and base URL, with Anthropic-compatible and OpenAI-compatible connection modes. Claude and DeepSeek endpoints are given as worked examples in the docs.
Everything is oriented around trust and autonomy — how much rope you give the agent, and how it proves what it did.
Shorthand for pulling a file into context, running a reusable command, or invoking a saved skill mid-task.
The agent can operate directly inside a remote host or container while the desktop app stays local — a genuine differentiator versus most cloud-only competitors.
Keeps objective state across a long task, recovering progress instead of restarting when a session breaks.
Multi-file feature work, debugging, refactors, test-writing, and review — anywhere a shell, git, or container workflow already exists. No published language whitelist.
The app itself is free. What you pay for is model access — either through Z.ai's GLM Coding Plan or your own BYOK provider bill.
● full support ◐ partial / conditional ○ not supported. Pricing shown is entry → top self-serve tier.
| Tool | Core LLM | Entry → top price | Agent capability | Offline | BYOK | Ease of use |
|---|---|---|---|---|---|---|
| ZCode | GLM-5.2 defaultGLM-5-Turbo option | Free appLite $16–18 → Max $144 | Agent-first ADEgoal mode, remote SSH/Docker | ◐remote-workspace focused | ●Claude, DeepSeek examples | Moderateagent-centric, new |
| Cursor | Multi-modelClaude, GPT, Gemini — no fixed default | Free HobbyUltra $200/mo | Strongbackground agents, Composer | ○ | ◐model choice, not a raw key | Easymost polished onboarding |
| VS Code | None built-inCopilot or extensions | FreeCopilot billed separately | Depends on extension | ●core editor is fully local | ●via extensions | Easyuniversal, huge ecosystem |
| Windsurf | SWE-1.6 defaultClaude / GPT selectable | FreeMax $200/mo | StrongCascade + cloud Devin agent | ○ | ◐Free/Pro only, not Teams/Enterprise | Easy–moderatenow rebranded Devin Desktop |
| Codex | GPT-5.x family | Bundled in ChatGPTPlus $20 → Pro 20x $200 | Strongcloud sandboxed tasks, CLI+web+IDE | ○CLI is local, execution calls the API | ●CLI supports raw API key | ModerateCLI/terminal-first |
| VSCodium | None built-in | Free | None nativelycommunity extensions add it | ●fully offline via Ollama/local models | ●fully open, any provider or local | Easy editorDIY for AI features |
| Antigravity | Gemini 3.x/3.5 defaultClaude Sonnet/Opus, GPT-OSS options | Free tierUltra $100–$249/mo | Strongmulti-agent Mission Control, browser subagent | ○offline on roadmap only | ●multi-model incl. Claude, GPT-OSS | Moderateambitious, reported quota instability |
ZCode is only days old, and the strongest claims about it need a second look before they go in front of anyone external.
"#2 on Code Arena" most likely refers to a human-preference, blind-vote leaderboard rather than a fixed benchmark suite. That's a real signal of developer preference in head-to-head comparisons, but it isn't proof of superiority on every task — worth stating precisely rather than as a bare ranking claim.
Reception evidence so far is launch-adjacent — creator walkthroughs and Z.ai's own docs — not a broad base of independent reviews. That's expected for a product days old; it just means the "widely praised" framing should wait for more independent coverage.