ZCode
Z.ai (Zhipu AI) · ZCode: Z.ai's Open-Source Agentic Development Environment for GLM-5.3
ZCode is Z.ai's free, open-source (Apache 2.0) desktop agentic development environment, built around its own ZCode Agent and tuned for the GLM-5.3 model. The agent plans, edits code, runs commands, and works through long multi-step goals, and you can follow and steer tasks from your phone or from WeChat and Feishu. It suits developers who want a GLM-first coding agent with a desktop workspace, especially those already on, or considering, a GLM Coding Plan.
Best for
Developers who want a desktop, agent-first coding environment tuned for GLM-5.3, particularly GLM Coding Plan subscribers who run long multi-step tasks and want to steer them from a phone or chat app
Not ideal for
Teams that need OS-level sandboxing out of the box, organizations that require a security review before code reaches a vendor's model service or gateway, and developers who mainly want a terminal-only agent or a model-neutral tool
Who it's for
Individual developers and small teams who want an agent-first desktop coding environment, especially users of Z.ai's GLM models and GLM Coding Plan
Capabilities
- ZCode Agent, Z.ai's first-party coding agent tuned for GLM-5.3, keeping file references, terminal output, Git state, and browser context in one task
- Four execution modes: Ask before changes (default), Edit automatically, Plan, and Full access, with Allow, Always Allow, Reject, and Always Reject decisions on gated actions
- Goal Mode (`/goal`): sets a session objective, runs a separate verification after every round, and keeps starting new rounds until the goal is met, with pause, resume, replace, and clear commands
- Dynamic workflows (v3.14.0+): one script orchestrates several subagents on a complex task, started from the input menu or `/workflow`, with concurrency adjustable while it runs
- Built-in general-purpose and read-only Explore subagents, plus user-defined custom subagents (beta)
- Remote development in SSH hosts, WSL distributions, and running Docker containers, where file access, commands, Git, and the agent all execute in the target environment
- Mobile Remote Control via QR code, and Bot Channel integrations for WeChat and Feishu to check progress and send instructions
- Scheduled tasks, plus idle-time tasks that personal Coding Plan subscribers can queue to run free during spare capacity (rolling out)
- Extensible with plugins, skills, MCP servers, custom commands, and lifecycle hooks, plus optional per-project long-term memory stored locally
- Model connections through a Z.ai or BigModel account and GLM Coding Plan (including team plans), an API key, or custom providers using the Anthropic or OpenAI protocols
- Desktop builds for macOS (Apple Silicon and Intel), Windows (x64 and ARM64), and Linux (x64 and ARM64, beta) as .deb, .rpm, and AppImage
- Apache-2.0 source on GitHub for the desktop client, web client, backend services, TUI, and agent CLI, with a NOTICE file on data flows, permissions, and local storage
Limitations
- Tuned for GLM-5.3. Other models connect through API keys or custom providers, but custom models default to a 200K context window until you change it, and idle-time tasks require a personal GLM Coding Plan rather than an API key
- In September 2026 Zhipu said a default-on code-indexing feature could upload repository data for cloud Repo Wiki generation, and that v3.14.0 removed that path. The NOTICE file also says requests to the official Coding Plan Anthropic-compatible endpoints are rerouted through a ZCode gateway with bodies and auth headers intact. Review it before using ZCode on sensitive code
- According to the NOTICE file, the shared agent executor has no default operating-system sandbox, so tools run with the permissions of the account running them, and the standalone CLI defaults to the unrestricted `yolo` mode for non-interactive `--prompt` runs unless `--mode` is set
- Questions the agent asks you (not permission requests or plan approvals) continue on the agent's own judgment after five minutes by default, unless you turn that setting off
- GLM Coding Plan usage is metered in credits with both 5-hour and weekly caps, and GLM-5.3 output costs 24 credits per 10,000 tokens at standard rates, so long agent sessions can hit the caps, especially on Lite
- The website distributes only the desktop app. The terminal (TUI) and web modes in the repository have to be built from source, and the Linux desktop build is labeled beta
- The Bot Channel docs cover WeChat and Feishu only, with DingTalk, Discord, and WeCom listed as coming later, and remote development needs a POSIX shell, so native Windows remote hosts are not supported
- Model input and output logs are written locally by default and may contain prompts and code, and stored command output in `~/.zcode/cli/exec` is not cleaned up automatically
Use cases
- Refactoring a module with Goal Mode set to "keep `pnpm test` passing" so the agent iterates until the check is met
- Running a large migration as a dynamic workflow that splits the work across several subagents
- Developing inside a remote Linux server, GPU box, WSL distribution, or Docker container from the desktop app
- Checking on and redirecting a long-running task from a phone or a WeChat or Feishu chat
- Queuing non-urgent work such as codebase analysis or comment backfills as idle-time tasks on a Coding Plan
- Using GLM-5.3 on a GLM Coding Plan in a desktop agent instead of a terminal tool
Our take
ZCode is Z.ai's case for using GLM-5.3 in a full desktop agent environment rather than through someone else's CLI. Goal Mode's per-round verification, subagent workflows, and remote development are well suited to long tasks, and the free app costs nothing extra if you already pay for a GLM Coding Plan. The same GLM focus is also its main limit, and trust needs attention: Zhipu open-sourced the client in response to its September 2026 upload episode. The NOTICE file that came with it is detailed, and it discloses no default OS sandbox, gateway rerouting of Coding Plan requests, and a `yolo` default for non-interactive CLI runs. Read it, keep the agent on Ask or Plan mode for unfamiliar code, and judge ZCode on your own repositories.
Who should use it
Developers and small teams who use or plan to use GLM-5.3, want a desktop workspace where an agent can work through long goals with verification, and value remote development plus the option to check on tasks from a phone, WeChat, or Feishu.
Who should skip it
Teams handling code that cannot leave controlled environments without a formal review, anyone who needs an OS-level sandbox by default, and developers who want a terminal-first or strictly model-neutral agent.
Strengths
- The app is free and open source under Apache 2.0, and the NOTICE file documents its data flows and permission model in detail
- Goal Mode verifies progress after every round, which suits long tasks that would otherwise need repeated "continue" prompts
- Subagent workflows, remote development, and phone or chat-app control are built in
- Uses GLM Coding Plan quota directly, and subscribers get free idle-time tasks as that feature rolls out
- Desktop builds for macOS, Windows, and Linux, including ARM64
Weaknesses
- ZCode Agent is tuned for GLM-5.3; other models work but are not the focus
- The September 2026 repository-upload episode means extra due diligence for sensitive code
- No default OS-level sandbox for agent commands
- Credit-based plan limits on 5-hour and weekly windows can be hard to predict for long tasks
ZCode pricing
ZCode app
Free
- Desktop app and Apache-2.0 source at no cost
- Bring a GLM Coding Plan, an API key, or a compatible third-party provider
- New users get a 5-day trial with a daily GLM token allowance
GLM Coding Lite
$18
Billed monthly
- 2,000 credits per 5 hours and 10,000 credits per week
- GLM-5.3 and GLM-5.3-Flash
- Supports 20+ agent tools, including ZCode and Claude Code
- Vision, Web Search, Web Reader, and Zread MCP tools
GLM Coding Pro
$80
Billed monthly
- 6x Lite usage (12,000 credits per 5 hours, 60,000 per week)
- Priority access to new flagship models and features
- Faster generation speeds
GLM Coding Max
$168
Billed monthly
- 14x Lite usage (28,000 credits per 5 hours, 140,000 per week)
- First access to new flagship models and features
- Dedicated resources during peak times
Free tier limits: The ZCode app is free. New users get a 5-day trial with a daily GLM token allowance that expires after the five days. It is not an ongoing free tier.
Note: Plan prices are the list prices shown on the ZCode homepage, and Z.ai's Coding Plan docs say plans start at $18/month. On September 30, 2026 the homepage showed discounted monthly prices of $12.60 (Lite), $56 (Pro), and $117.60 (Max). In China the plans are sold through BigModel in yuan. Usage is metered in credits: GLM-5.3 costs 6.9 credits per 10,000 input tokens, 1.7 per 10,000 cached input tokens, and 24 per 10,000 output tokens, and off-peak usage is charged at 50% (peak hours are 14:00–18:00 UTC+8 on weekdays). Monthly, quarterly, and yearly billing are available in the app. Idle-time tasks run free for personal Coding Plan subscribers as that feature rolls out. Using your own API key bills through that provider instead.
Technical specs
Context window
1M tokens with GLM-5.3 (custom models default to 200K unless configured)
Available models
Where ZCode excels
Getting a failing test suite back to green
Goal Mode checks after every round whether the stated objective, such as all tests passing, has been met, and starts another round only if it has not, so you do not have to keep prompting the agent to continue.
Working on a project that only runs on a remote Linux box
Remote development runs file access, commands, Git, and the agent on the SSH host, WSL distribution, or container, while the desktop app keeps the interface and model configuration local.
Following a long task while away from the desk
Remote Control opens the desktop window's tasks on a phone, and a WeChat or Feishu bot can report progress and accept new instructions for the same workspace.
ZCode vs. competitors
ZCode vs. Claude Code
Claude Code is Anthropic's agent for the terminal, IDEs, and desktop, built around Claude models and Anthropic subscriptions. ZCode is a desktop-first ADE built around Z.ai's own agent and GLM-5.3, with Goal Mode, dynamic subagent workflows, and WeChat and Feishu bots. GLM Coding Plans also work inside Claude Code, so the choice is partly about which agent you prefer to drive GLM.
ZCode vs. Kimi Code
Kimi Code is Moonshot AI's agent for its Kimi models, available as an open-source CLI, a desktop app, and IDE integrations under a Kimi membership. ZCode is Z.ai's counterpart for GLM-5.3 and centers on the desktop app, with Goal Mode, remote development, and chat-app control. Each is tuned for its own lab's models.
ZCode vs. Qwen Code
Qwen Code is Alibaba's Apache-2.0 agent for the terminal, desktop, and IDEs, designed to work with many model providers. ZCode is also Apache-2.0 but is tuned first for GLM-5.3 and the GLM Coding Plan, with Goal Mode verification and idle-time tasks for subscribers.
ZCode vs. Google Antigravity
Google Antigravity is Google's agent-first platform with a desktop command center for parallel agents, an agentic IDE, a CLI, and an SDK, free for individuals with higher limits on Google AI plans. ZCode is a single desktop ADE around one first-party agent and GLM-5.3, with open-source code and GLM Coding Plan billing.
Frequently asked questions
What is ZCode?
ZCode is Z.ai's (Zhipu AI's) agentic development environment. It is a desktop app built around ZCode Agent, a first-party coding agent tuned for GLM-5.3, with a file tree, terminal, Git tools, a built-in browser, Goal Mode for long tasks, subagent workflows, remote development, and remote control from a phone or from WeChat and Feishu.
Is ZCode free?
The app is free, but you need model access: a Z.ai or BigModel account with a GLM Coding Plan, an API key, or a compatible third-party provider. New users get a 5-day trial with a daily GLM token allowance. The ZCode homepage lists the Coding Plan at $18/month for Lite, $80/month for Pro, and $168/month for Max, and on September 30, 2026 it showed discounted prices of $12.60, $56, and $117.60.
Is ZCode open source?
Yes. Since September 2026 the source has been on GitHub at zai-org/ZCode under the Apache 2.0 license. It includes the Electron desktop client, the web client, backend services, and the agent CLI and TUI. Third-party components keep their own licenses. The NOTICE file says the public release does not promise every feature of the official product, and that the Computer Use package included there is a non-functional placeholder.
What happened with ZCode and repository uploads?
In mid-September 2026 users questioned whether ZCode was uploading repository data. According to IT Home and Tencent News, Zhipu said a code-indexing feature, on by default at launch, could upload repository data when Repo Wiki pages were generated in the cloud. It said the uploaded data was deleted after generation and never used for training, and it apologized. On September 21 it announced that v3.14.0 had removed the Repo Wiki feature and the local repository snapshot upload path, and that the client was open source. It also cited audits by CAICT and NSFOCUS finding that the cloud storage bucket involved held no data and had been deleted. ZCode's own v3.14.0 changelog lists a fix for abnormal uploads in the repository wiki. The current docs describe a repository Wiki that stores its output locally and sends filtered code context to the model service you select.
Can I use models other than GLM in ZCode?
Yes. Besides Z.ai and BigModel accounts, ZCode accepts API keys and custom providers that use the Anthropic or OpenAI protocols, including self-hosted services. The docs walk through Anthropic, OpenRouter, Kimi, and DeepSeek setups. ZCode Agent is tuned for GLM-5.3, though, and custom models default to a 200K context window unless you change it.
Does ZCode sandbox the agent?
Not at the operating-system level by default. ZCode gates file edits, commands, and other sensitive actions through its execution modes and approval prompts, but the repository's NOTICE file says the shared agent executor provides no default OS sandbox and that tools run with the permissions of the account running them. For untrusted projects, the NOTICE recommends limiting the account, credentials, and network access; working inside a remote container is another option.
Which platforms does ZCode support?
The desktop app runs on macOS (Apple Silicon and Intel), Windows (x64 and ARM64), and Linux (x64 and ARM64, currently beta) as .deb, .rpm, or AppImage. Workspaces can also live on SSH hosts, in WSL on Windows, or in running Docker containers.
ZCode vs Claude Code: what is the difference?
Claude Code is Anthropic's coding agent for the terminal, IDEs, and desktop, built around Claude models. ZCode is a desktop-first environment built around Z.ai's own agent and GLM-5.3, with Goal Mode, subagent workflows, and phone or chat-app control, and it can import past Claude Code conversations. GLM Coding Plans can also be used inside Claude Code itself.
Integrations & fit
Alternatives to consider
Anthropic
Developers and teams who want deep codebase reasoning, multi-file changes, and persistent project instructions in a terminal-friendly or IDE-native workflow
Paid
Moonshot AI
Developers who want Moonshot's Kimi K3 models in one coding agent across the terminal, a desktop app, VS Code, and ACP editors, paid for through a Kimi membership
Paid
Alibaba (Qwen team)
Developers and teams who want a free, open-source, Claude Code-style coding agent that works with Qwen models or any major provider, especially teams already on Alibaba Cloud or using DingTalk, Feishu, or WeChat
Free
Developers who want to supervise several autonomous coding agents in parallel from a dedicated desktop app, with a free tier and access to both Gemini and Claude models
FreemiumAbout ZCode
ZCode calls itself an Agentic Development Environment (ADE): instead of an editor with an assistant attached, the desktop app centers on ZCode Agent, Z.ai's first-party agent, with a file tree, terminal, Git tools, and a built-in browser preview around it. You describe a task, pick one of four execution modes (ask before changes, edit automatically, plan first, or full access), and the agent reads the workspace, changes files, runs commands, and surfaces each change for review. For longer work, Goal Mode (`/goal`) sets an objective and runs a separate check after every round, starting another round until the goal is met. Dynamic workflows, added in v3.14.0, let one script coordinate several subagents on a large task. Workspaces can be local folders, SSH hosts, WSL distributions, or Docker containers, and tasks can be followed from a phone through Remote Control or from a WeChat or Feishu bot. The app itself is free. Models come from a Z.ai or BigModel account (GLM Coding Plan, with a 5-day trial for new users), an API key, or any provider that speaks the Anthropic or OpenAI protocols. The product is plainly tuned for GLM-5.3 and its 1M-token context, though, and some benefits, such as free idle-time tasks, require a Coding Plan. Trust is the main thing to weigh. In September 2026 users questioned whether ZCode was uploading repository data. According to Chinese tech press, Zhipu traced it to a code-indexing feature that was on by default and could upload repository data when Repo Wiki pages were generated in the cloud, apologized, and said v3.14.0 removed that upload path. It then open-sourced the client on GitHub under Apache 2.0. The repository's NOTICE file documents the client's outbound requests, gateway routing, and local logs, and states that the agent executor has no default operating-system sandbox. ZCode fits developers who are willing to read those disclosures and run the agent with appropriately limited permissions.
Updates from ZCode
Zhipu announced that ZCode's source is public on GitHub (zai-org/ZCode) under the Apache 2.0 license, covering the desktop client, web client, backend services, TUI, and agent CLI. The announcement followed its apology to users over repository data uploads linked to the Repo Wiki feature.
v3.14.0 added dynamic workflows that coordinate multiple subagents from one script, switching between Office and Coding modes, custom keyboard shortcuts, and an onboarding guide. Its changelog also lists a fix for abnormal uploads in the repository wiki.
Are you the founder? Claim this listing →