ZCode
Z.ai (Zhipu) · ai coding agents · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-11
Z.ai's open-source coding agent (Apache-2.0) — desktop + browser + CLI, tuned for GLM-5.3/Flash, with sub-agents, plan mode and one of the widest provider menus in the field.
#8 of 15 in AI Coding Agents
ToolScore breakdown — why 75.5/100
10 criteria × 10 points. Low pillars are visible on purpose — that's the point of a score you can audit. Methodology →
What it does well
- ▸One of the best open-source coding agents right now: clean, intuitive management menus, plan mode, and sub-agents that work perfectly
- ▸Tuned for GLM-5.3 / GLM-5.3-Flash — excellent quality-per-token when paired with them; bring a powerful LLM and it shines
- ▸Massive preset provider menu (Kimi Code, MiniMax M Plan, DeepSeek, Alibaba, Tencent, Volcano, OpenRouter, OpenCode Go/Zen...) + custom base-URL providers incl. local servers like Ollama — config takes seconds
- ▸Fully open-source client (Apache-2.0) since Sept 2026 — auditable and buildable; desktop + browser + CLI with day-one Linux support
Where it costs you
- ▸The harness is optimized for GLM models — behavior with other models can vary; with a weak LLM the agent shows it
- ▸If you fuel it with the GLM Coding Plan, Z.ai's terms apply (training on individual plans, data processed in China) — the plan's terms, not the agent's
- ▸Custom request parameters are silently ignored (hand-edited config values never reach the request)
- ▸Young product moving fast; the Sept 2026 privacy incident is fixed and open-sourced, but the history exists
Hidden traps — the stuff the pricing page doesn't say
Training on your code (individual plans) and China data residency are Z.ai Coding Plan / Z.ai API terms. Use another provider or a local server and none of it touches your code — that freedom is the point of the open client.
The client once silently uploaded .git workspaces to Aliyun OSS. It was fixed within days and the client was open-sourced (Apache-2.0) — now auditable and buildable. Relevant caution if you route through Z.ai; irrelevant with your own provider.
Expect the best experience with GLM-5.3/Flash. Other models work via custom providers, but the harness's prompt paths and defaults assume GLM behavior.
Values like reasoning_effort added by hand to config.json are silently dropped — never written into the request body.
Know the limits before you pay: read the traps above, then decide with eyes open.
Visit official siteAffiliate link where available — never affects the score.
- official ZCode official site verified
- official GLM Coding Plan docs (credits, limits) verified
- official GitHub zai-org/ZCode (Apache-2.0, opened Sept 2026) verified
- benchmark Ferstar: silent workspace upload (security research) verified
- community HN: ZCode harness thread (511 pts) verified
- official ZCode docs: configuration (custom providers / BYOK) verified
Data as of 2026-11-11. Prices and limits change monthly — report anything stale. How we verify: methodology.
Closest alternatives
Open-source VS Code agent with BYOK transparency — powerful, but famous for eating tokens by design.
Open-source VS Code/JetBrains/CLI agent — a merge of Roo Code and Cline, acquired by Anaconda in Jul
Open-source autonomous agent (ex-OpenDevin) — every conversation in its own Docker sandbox, 93k star
ZCode review
In-depth analysis · last updated 2026-11-11
The agent, judged as an agent
Strip away the vendor story and ZCode is simply one of the best open-source coding agents you can run today. The management UI is clean and intuitive — providers, models and settings are a few clicks, not config archaeology. Plan mode for architecture-first work, sub-agents that delegate cleanly, and the same engine in desktop, browser and CLI with day-one Linux support. Fire a hard task at it with a capable model and it plans, executes and reviews like the top tier of this hub.
Tuned for GLM — but not locked to it
ZCode is built around GLM-5.3 and GLM-5.3-Flash, and that pairing is where it performs best — excellent quality-per-token, "worlds cheaper" than Opus per HN users, and a planner that never refuses a task. But the provider menu is one of the widest in the field: ~30 presets (Kimi Code, MiniMax M Plan, DeepSeek, Alibaba Coding Plan, Xiaomi, Tencent TokenHub, Volcano Ark, StepFun, OpenAI, Anthropic, xAI, OpenRouter, OpenCode Go/Zen) plus a fully custom provider — any base URL, Anthropic or OpenAI format, your key, your model list. That includes local OpenAI-compatible servers like Ollama. One caveat both ways: the harness assumes GLM behavior, so other models can vary — and a weak LLM will show, because the agent is only as strong as the brain you give it.
The plan's terms are not the agent's
If you fuel it with the GLM Coding Plan ($18/mo Lite, credits-based), Z.ai's terms ride along: training on your code on individual plans, data processed in China, Beijing-time peak windows, no refunds. September's incident — the client silently uploading .git workspaces — happened, was fixed within days, and the client was open-sourced (the repo appeared three days after the story broke). We keep that history on the page scoped where it belongs: it's a caution about Z.ai endpoints, not about the open client. Point it at your own API key or a local model and none of it touches your code. Full plan terms live in our coding plans ranking.
Verdict
75.5/100: a top-tier open agent — intuitive, sub-agent capable, GLM-tuned but genuinely provider-agnostic. Pair it with a strong LLM (GLM for value, or whatever you trust) and it competes with anything in this hub. Just read the fuel's terms, not only the agent's.