Published Updated
In one or two sentences, before anything else.
ZCode is Zhipu AI's (Z.ai) free desktop coding-agent app for macOS, Windows and Linux, built around its GLM-5.3 model, with model use billed through the GLM Coding Plan from 18 USD a month. First released as Z Code in December 2025, it switched to Zhipu's own agent in June 2026; Z.ai said in August it had passed one million users.
What it can do, one sentence each, and whether we actually tested it.
Splits a larger goal into planning, execution and verification and keeps working through the steps, per Z.ai's ZCode page.
Z.ai's description. In ifanr's August 2026 test one 3D build ran for more than an hour.
Ask before every change (the default), edit automatically, plan first, or full access with few confirmations; Shift+Tab cycles through them.
Per ZCode's docs. Its terms make you responsible for what the agent does, full access included.
Starts and steers tasks from WeChat, Feishu or Telegram, or mirrors a session to a phone browser.
ZCode's data notice says session content passes through an external relay and the chat platform can read bot messages.
Signing in with a BigModel or Z.ai account gives GLM-5.3 and GLM-5.3-Flash on the GLM Coding Plan.
Zhipu says plan subscribers get 150% of their quota inside ZCode.
Connects to Anthropic, OpenAI, OpenRouter, Moonshot Kimi, MiniMax, Xiaomi MiMo or any Anthropic- or OpenAI-compatible endpoint with your own key.
Generates a wiki of your repository, runs subagents, and supports MCP servers, plugins, skills and hooks.
Per ZCode's docs. The wiki sends filtered code context to the model you choose and is stored locally.
Lets subscribers queue non-urgent work that runs in spare capacity without using plan quota.
Rolling out gradually, per ZCode's docs.
Publishes the desktop app, web client and command-line agent on GitHub under the Apache-2.0 licence (zai-org/ZCode).
eesel described ZCode as closed source in July 2026.
The people it helps and the problems it solves, and where it ranks.
Built for
Problems it solves
Graded in
Public tests other people ran: what they gave it and how it did.
Case 1
Input
A 3D web simulation of two planets colliding, with crust breaking up and debris (prompt not published in full).
What happened
ZCode kept testing and fixing its own page for more than an hour, longer than any task the reviewer had run in Claude Code. The result showed cracking crust, lava and a debris ring. One run, and a long one.
Case 2
Input
使用 Three.js 创建大疆 Mavic 3 Pro 无人机的 3D 模型 (build a Three.js 3D model of a DJI Mavic 3 Pro drone), then a watch replica from an uploaded photo.
What happened
The drone was done in under two minutes with adjustable propellers and camera. The watch took over ten minutes and its screen looked like it floated in the air; ZCode fixed it after one follow-up.
Case 3
Input
Review the latest post in a repository of more than 1,700 posts and suggest up to four improvements, with GLM-5.3 at low effort and "Ask before changes" on.
What happened
It reported dead-end links, a long section without a table of contents and a timezone inconsistency, and it asked before running a shell command. Copes called it a useful second tool, not a replacement for his main setup.
What it does badly, with the evidence, and what we could not settle.
Weak
There is no code editor.
Evidence: Flavio Copes (updated October 6, 2026) notes ZCode omits a traditional editor and calls it a useful second tool, not a replacement for his main setup.
Free use is for non-commercial purposes only.
Evidence: ZCode's terms (effective June 15, 2026) limit use to non-commercial personal research and study unless you subscribe or have a separate agreement.
You carry the risk of what the agent does.
Evidence: The terms treat the agent's actions as an extension of your own, tell you to back up data and use a sandbox, and cap ZCode's liability at the greater of six months' payments or 100 USD.
Shared sessions are public by default.
Evidence: ZCode's data notice on GitHub says a shared session includes your text, model replies, reasoning and tool input and output, and the sharing screen defaults to publicly importable.
Long runs and quota use.
Evidence: ifanr's planet test ran for more than an hour. eesel relayed launch-week reports (July 2026) that it was about half as fast as Claude Opus 4.8 and used quota faster.
Not for code that may not leave your machine.
Evidence: Prompts, code, diffs and attachments go to the model service you pick, and official plans route through ZCode's gateway (data notice on GitHub). Copes would keep client work off it until data rules are checked; teams under such rules should wait.
Uncertain
How much the free trial gives.
Why unsure: ZCode's docs give new BigModel users five days, but the welcome page lists 3M GLM-5.3 tokens a day plus 2M for GLM-5-Turbo, the model page 3M plus 5M for GLM-5.3-Flash, and a Z.ai screenshot 3M a day.
What one million users means.
Why unsure: Z.ai's August 2026 figure, as reported by RuntimeWire, does not say whether it counts registered, active or paying users.
GLM-5.3's lead on coding benchmarks.
Why unsure: The 50% gain over GLM-5.2 is measured on Z.ai Code Bench, Z.ai's private benchmark with no published task count (DataNorth, August 17, 2026).
The numbers, each with its source and the date we checked it.
| Spec | Value | Note and source |
|---|---|---|
| Maker | Beijing Zhipu Huazhang Technology Co., Ltd. (Zhipu AI, Z.ai outside China) | Source: zcode.z.aichecked |
| First release | December 26, 2025, as Z Code (alpha, Mac and Windows) | Source: aibase.comchecked |
| ZCode 3.0, own agent replaces third party agents | June 13, 2026 | Source: news.17173.comchecked |
| Current version | 3.15.3, October 13, 2026 | Source: zcode.z.aichecked |
| Platforms | macOS (Apple Silicon, Intel), Windows (x64, ARM64), Linux (x64, ARM64, beta) | Source: zcode.z.aichecked |
| App price and trial | Free app; new BigModel users get a 5-day model trial; otherwise models come from a GLM Coding Plan or your own API key | Source: zcode.z.aichecked |
| GLM Coding Plan on Z.ai, billed monthly | Lite 18, Pro 80, Max 168 USD a month; 20% off quarterly, 30% off yearly | Source: z.aichecked |
| GLM Coding Plan on BigModel, billed monthly | Lite 118, Pro 538, Max 1,078 CNY a month; 20% off quarterly, 30% off yearly | Source: bigmodel.cnchecked |
| Usage limits | Lite 2,000 credits per 5 hours and 10,000 a week; Pro 6 times and Max 14 times Lite; half the credits off-peak | Source: docs.z.aichecked |
| Training on your data | Off by default (opt-in); international service run from Singapore by JINGSHENG HENGXING TECHNOLOGY PTE. LTD. | Source: zcode.z.aichecked |
| Website visits, September 2026 (zcode.z.ai) | 3.07 million (+27.2% on August); 20th of 20 on AICPB's AI tools in China list; 1st of 10 on AICPB's China growth list | Source: mp.weixin.qq.comchecked |
Answer first, then why, an example, the limits and what to do next.
The ZCode app is free, but the GLM models it runs need a GLM Coding Plan or your own API key. The plan starts at 18 USD a month on Z.ai (12.6 USD a month billed yearly) or 118 CNY on BigModel, and new BigModel users get a five-day trial.
Limits
ZCode's terms limit unpaid use to non-commercial personal research and study.
ZCode is made by Zhipu AI, known as Z.ai outside China, and it is the same product as Z Code. Zhipu released Z Code on December 26, 2025 as a desktop front end for Claude Code, Codex and Gemini; ZCode 3.0 replaced them with Zhipu's own agent in June 2026.
ZCode runs GLM-5.3 and GLM-5.3-Flash by default through a BigModel or Z.ai account. You can also connect Anthropic, OpenAI, OpenRouter, Moonshot Kimi, MiniMax, Xiaomi MiMo or any Anthropic- or OpenAI-compatible endpoint with your own key.
No public test has compared them on the same task, so nobody can say yet. ZCode is a desktop agent app without a code editor; Claude Code runs in the terminal and comes with Claude Pro at 20 USD a month; Cursor is a code editor whose cheapest paid plan, 8 USD, covers Grok models only.
Next step
The GLM Coding Plan also works inside Claude Code, Cursor and TraeCode, so you can try GLM-5.3 without switching tools.
ZCode sends code to the model service you choose, but it does not use your data for training unless you opt in. Its privacy policy keeps the Optimization Program off by default. Its data notice on GitHub says prompts, code, diffs and attachments go to the model endpoint, official plans pass through ZCode's gateway, and shared sessions are public by default.
Yes, ZCode's source code is on GitHub under the Apache-2.0 licence, in the zai-org/ZCode repository. It holds the desktop app, a web client and the command-line agent. eesel described ZCode as closed source in July 2026, so it was opened later; we could not find the date.
This page is built from what can be checked today: Z.ai's own site, docs, changelog, terms and GitHub repository, news reports from December 2025 to August 2026, and three hands-on reports by other people. Every price and date links to where it came from, and Z.ai's claims are labelled as Z.ai's.
According to AICPB's traffic rankings, zcode.z.ai had 3.07 million visits in September 2026, up 27.2% on August: 20th among AI tools in China and 1st on the fastest-growing list in China.
ZCode began as a desktop shell for other companies' coding agents and became Zhipu's own GLM agent.
| Date | Event | Source |
|---|---|---|
| December 26, 2025 | Zhipu releases Z Code, an alpha desktop app for Mac and Windows that runs Claude Code, Codex and Gemini with one API key | AIbase, IT Home via ifeng |
| June 13, 2026 | ZCode 3.0 switches to Zhipu's own ZCode Agent, stops maintaining other agents and is tuned for GLM-5.2 | 17173 |
| July 1, 2026 | Z.ai launches ZCode as a free desktop coding environment built around GLM-5.2 | RuntimeWire, citing VentureBeat |
| August 2026 | Z.ai says ZCode has reached one million users and resets GLM Coding Plan limits | RuntimeWire, August 11 |
| August 14, 2026 | GLM-5.3 arrives through ZCode and the GLM Coding Plan before its public API | DataNorth |
| September 19, 2026 | Version 3.14.0 adds dynamic workflows | changelog |
| October 13, 2026 | Version 3.15.3, the current release | changelog |
The source code, covering the desktop app, a web client and a command-line agent, is now on GitHub under the Apache-2.0 licence. eesel described ZCode as closed source in July 2026; we could not find the date it was opened.
The app is free. The models are not: ZCode signs in to a BigModel (China) or Z.ai (international) account and uses its GLM Coding Plan, or an API key you bring.
| Plan | Z.ai, billed monthly | Z.ai, billed yearly | BigModel, billed monthly | Usage |
|---|---|---|---|---|
| Lite | 18 USD | 12.6 USD a month | 118 CNY | 2,000 credits per 5 hours, 10,000 a week |
| Pro | 80 USD | 56 USD a month | 538 CNY | 12,000 per 5 hours, 60,000 a week |
| Max | 168 USD | 117.6 USD a month | 1,078 CNY | 28,000 per 5 hours, 140,000 a week |
Sources: Z.ai, BigModel and Z.ai's plan docs. Quarterly billing is 20% off. Outside peak hours (14:00 to 18:00 UTC+8 on weekdays) model use costs half the credits. Zhipu says subscribers get 150% of their quota inside ZCode. New BigModel users get a five-day trial with no payment.
Three hands-on reports, none of them a controlled benchmark:
GLM-5.3's headline gains, such as 50% over GLM-5.2 on Z.ai Code Bench, are Z.ai's own figures on a private benchmark.
| ZCode | Claude Code | Cursor | TRAE | |
|---|---|---|---|---|
| Form | Desktop agent app, no code editor | Runs in the terminal | Code editor | TraeCode coding agent and TraeWork assistant |
| Cheapest way in | Free app; GLM plan from 18 USD a month | Claude Pro, 20 USD a month | Free plan; paid from 8 USD (Grok models only) | 4,500 credits for new users |
| GLM-5.3 | Built in | With a GLM Coding Plan | With a GLM Coding Plan | With a GLM Coding Plan |
Sources: Flavio Copes, Anthropic, Cursor, TRAE, TRAE China and Z.ai. No public test we found runs the same task in all four.
The GLM Coding Plan lists Claude Code, Cursor, TraeCode, TraeWork, Codex and OpenCode among its supported tools, so GLM-5.3 does not require ZCode. What ZCode adds is the extra quota, idle-time tasks that run without using quota, and a harness Zhipu tunes for its own models.
Sources
Changelog
T3 Code is a free, MIT-licensed app from T3 Tools Inc., credited to Theo Browne's team, that runs the coding agents you already pay for, such as Claude Code, Codex, Cursor and OpenCode, side by side from a desktop, web or phone interface. First released on March 7, 2026, its GitHub repository had 28.1k stars on October 17, 2026.
Command Code is a terminal coding agent from Langbase, Inc., trading as Command Code and led by Ahmad Awais, that says it learns your coding "taste" from the changes you accept, reject and edit. Announced with 5 million USD in seed funding on February 25, 2026, it now focuses on low-cost open models, with plans from 1 USD a month.
VideoEnhancer is a web tool that upscales short MP4 clips to 1080p, 2K or 4K with the FlashVSR model and also compresses, face-swaps and generates video. After a 5-second free trial, 1080p costs 3 credits a second, $0.03 to $0.09 depending on how credits are bought. VideoEnhancer is our own product: it is made by the team behind GenGrades.
LayerGrab splits a flat image into up to 16 named, transparent layers over a filled-in background, on the web, in Figma or in Chrome, or pulls out only what you name. Each layer costs $0.10 on its $12 plan, $0.08 on Premium. No independent review exists yet. LayerGrab is our own product: it is made by the team behind GenGrades.
VideoTextRemover is a web tool that erases burned-in subtitles, captions and other on-screen text from MP4 and MOV videos and repaints what was behind them, up to 1080p. It bills per second processed, about $0.025 to $0.049, with 30 free seconds a month in some countries. VideoTextRemover is our own product: it is made by the team behind GenGrades.
MagicRemover is a browser tool that erases people, objects, text, logos and watermarks from photos, and subtitles from videos. Photos get 2 free removals a day, 2 more after sign-in, at about 1 megapixel; beyond that, credits start at $9.90 for 330. MagicRemover is our own product: it is made by the team behind GenGrades.