ZCode
The official shell for GLM-5.3; plans are billed in weekly credits.
Verified: Sep 12last change: Sep 18next price check: Dec 11
Facts with sources
- The system prompt is almost entirely environment restrictions: don’t start the dev server, don’t run a build. Source (in Russian)
- Claude Code + GLM-5.3 on Terminal-Bench: 41.82% of tasks solved for $2.7K, versus 51.82% for $6.0K with Opus 5. Source (in Russian)
- Z.ai access isn’t restricted from Russia. Source (in Russian)
Strengths
- Metered per prompt, not per token: Lite gives about 80 prompts per 5 hours and 400 a week, however many files the agent edits inside one.
- Off-peak calls (anything outside 14:00–18:00 UTC+8 on weekdays) cost half the points, which covers almost the whole US and EU workday.
- The app keeps the GLM cache above 98%; AICodeKing puts that at roughly 30% more effective tokens than a third-party harness.
- Goal mode, project indexing, skills, MCP, a built-in browser, phone remote control and Telegram, WeChat and Feishu bots are included.
Weaknesses
- Subagents are weak: project memory skips them, 3 of Superbash's 16 projects would not run, and 40% of a weekly Lite limit went in a day.
- GLM-5.3 cannot see images; the screenshot MCP is only on Pro and above, so on Lite you cannot show it a broken screen.
- Reasoning effort defaults to max, so a one-line fix burns quota on overthinking unless you lower it or say 'think less'.
- Past roughly 200k tokens in one thread the output visibly degrades; you have to open a new thread and re-summarise.
Example from practice
An uptime-monitoring MVP on the free tier
Shcoder used Plan mode with GLM-5.3 at high reasoning effort to design a service that pings sites every minute and reports uptime as percentages with charts, then handed the build to the cheaper GLM-5.3 Flash. The service worked on the first pass: sites could be added, checks ran, uptime was computed. The agent skipped the incidents section the first time and had to be asked again; the dark theme was fine, the light theme came out rough. He deployed it publicly; the video does not state how many free-tier tokens the project consumed.
source: Shcoder, '100 МИЛЛИОНОВ токенов бесплатно?! Тестирую новый GLM 5.3 в ZCode' (in Russian)
What reviewers say
Reviewer opinions on their own tasks and hardware, not measurements. Links go to the videos; transcripts live in docs/raw-data.
- GLM-5.3 + Z-Code (New Free Tier & BEST WAYS): This is the CHEAPEST GOOD CODER RIGHT NOW! · AICodeKing, Aug 18In the reviewer's view, Lite at $12.60 with caching, off-peak hours and goal mode gives a Claude-class model, and he struggles to hit the Lite limits.
- GLM 5.3 Review (Don't Use Subagents on ZCode) · Superbash (BoxminingAI), Aug 17The reviewer rates the model highly but ZCode's subagents poorly: let GLM-5.3 write the code itself rather than delegate.
- 100 МИЛЛИОНОВ токенов бесплатно?! Тестирую новый GLM 5.3 в ZCode · Shcoder, Sep 12The reviewer recommends trying it: the free pool covers a small MVP, and he finds both the subscription and the API cheap.
- Best AI agent harnesses and how I use them (pi, omp, Zcode) · 0xSero, Sep 2The reviewer calls ZCode the desktop app closest to Codex, with loops, automations and computer use built in, and uses it daily.
Who it is for
- Team on Next.js that needs an inexpensive model with access from Russia
Models in this agent
Score is from our manager-task benchmark; models without a score are not measured yet.
What changed
How to brief an agent like this
The four-line brief is the Foundation module: prompt structure, LLM agents and MCP, an AI policy for the team. It starts with the free open module.