← Back to research
·12 min read·company

ZCode

Z.ai's ZCode is a first-party coding ADE with GLM-5.3, desktop and CLI surfaces, and local execution. Review permissions, Coding Plan pricing, and the September 2026 snapshot-upload incident.

https://zcode.z.ai/en ↗
coding-agentselectrongithubgitmcp

Key takeaways

  • ZCode Agent is Z.ai's self-developed default agent inside a desktop ADE, with GLM-5.3 as the primary model family and optional third-party model connections.
  • The application is free; model access uses a GLM Coding Plan from $18/month, a China BigModel plan, prepaid API usage, or a bring-your-own key.
  • Execution is local or remote SSH/WSL/Docker with permission prompts, not a documented VM sandbox.
  • A September 2026 investigation found silent workspace snapshot uploads; Z.ai later stripped that pipeline, published Apache-2.0 source, and left historical-data questions unanswered.

FAQ

What is ZCode?

ZCode is Z.ai's agentic development environment. ZCode Agent is the first-party default agent for planning, editing, commands, browser control, and review in a desktop app, with a separately documented command-line and web surface in the public repository.

Is ZCode free?

The ZCode application is free. You still need a model plan or API key. Global GLM Coding Plan documentation starts at $18/month; China users can subscribe through BigModel. A five-day trial is documented, with conflicting daily token figures across pages.

Does ZCode run locally?

The desktop client runs on macOS, Windows, and Linux. File edits and commands execute on the local machine or on a connected SSH, WSL, or Docker target. Model inference still goes to the connected provider.

Did ZCode upload repositories to the cloud?

A September 2026 first-hand analysis of client 3.12.3 reported silent workspace snapshot packaging and Aliyun OSS uploads. Z.ai said a Wiki-related indexing feature caused uploads, later removed the pipeline in 3.14.x, and published source. Historical cloud copies remain unverifiable from outside.

Executive Summary

ZCode is Z.ai's first-party coding agent harness. The product is an Agentic Development Environment: you describe work in natural language, and ZCode Agent (the default self-developed agent) plans, edits files, runs commands, drives a built-in browser, and keeps the task in one workspace. It is tuned for the GLM-5.3 family and can also connect Anthropic, OpenAI, OpenRouter, and other compatible endpoints.[1][2][3]

Status checked September 23, 2026: the current desktop release is v3.14.3 (September 22). Official install docs ship macOS, Windows, and Linux desktop builds. The public GitHub repository zai-org/ZCode also documents a zcode command-line/TUI/web surface and an Apache-2.0 license. That repository appeared on September 21 after a workspace-snapshot upload incident; it is a flattened dump, not a long public history of the client that was reverse-engineered days earlier.[4][5][6][7]

This profile is the missing individual page called out in Foundation Lab Coding Agents. It is not a claim that ZCode already belongs in that comparison's member set.

AttributeCurrent evidence
ProviderZ.ai / Zhipu GLM; terms name JINGSHENG HENGXING TECHNOLOGY PTE.LTD as the ZCode service provider.[8][2]
ExecutionLocal desktop process, or SSH / WSL / Docker remote workspace. Permission modes gate edits and commands. Not a documented VM sandbox.[9][10][8]
InterfacesDesktop ADE (documented installer); GitHub README also describes TUI, local web UI, and Agent CLI.[5][6]
Entry tierApp free; five-day GLM trial, then GLM Coding Plan from $18/month, China BigModel plans, or BYOK.[11][12][3]
CategoryFoundation Lab Coding Agents

Product Overview

ZCode Agent versus other harnesses

ZCode Agent is the default agent when you create a task, not a generic wrapper that only brokers GLM into Claude Code or OpenCode. Docs describe workspace awareness, @ file mentions, / commands, $ skills, AGENTS.md instructions, optional project memory, four execution modes, thought levels for GLM-5.3, side conversations, and conversation forks.[1]

The GLM Coding Plan is a separate subscription that can also be used in Claude Code, Cline, and OpenCode. Using those tools with a GLM key is not the same product as ZCode Agent. ZCode's own FAQ says the application is free and that you still need a plan or API key.[12][11]

A useful first evaluation is a bounded repository change with a test command:

  1. Install the desktop app for your OS and complete first-launch model connection (Z.ai globally, BigModel in China, or an API key).[5][3]
  2. Open the project workspace (or a remote SSH/WSL/Docker target if the code only exists there).[10]
  3. Start in Ask before changes or Plan mode. Default Ask before changes confirms every edit and command; Plan mode waits for a plan before implementation.[9]
  4. State the change, the files, and the verification command. Use @ for the relevant paths. This is an illustrative evaluation, not a task reproduced for this report.
  5. Review diffs and command output in the task UI, then commit through your normal Git process. ZCode's terms treat agent file and command operations as your actions.[8]

Surfaces

SurfaceRoleBoundary to understand
Desktop ADEOfficial installer path: Electron app with chat, files, terminal, browser, Git, and review.[5][2]Agent work runs with the user's local (or remote) privileges.
Remote workspaceSSH, Windows WSL, or local Docker: files and commands execute on the target; accounts and model config stay on the desktop.[10]Native Windows remotes are unsupported. MCP secrets copied to a remote host are written as-is.
Command-line / web (source repo)zcode TUI, zcode --web, and Agent CLI in apps/zcode-cli.[6]Documented in the public repository, not on the current install page's download list.
Mobile Remote / botsPhone control of an existing desktop session; Feishu / WeChat bots in product docs.[2][10]Remote Control cannot create a new SSH/WSL/Docker connection.

Subagents, skills, and browser

Built-in general-purpose and read-only Explore subagents can run in the foreground or background. Custom user-level subagents are labeled beta and live under ~/.zcode/agents/; workspace-level custom subagents are not available from Settings yet. Explore does not inject AGENTS.md.[13]

Browser control uses an official Browser Use plugin (on by default) so the agent can open pages, click, and screenshot after frontend changes.[1] MCP servers are supported; Coding Plan docs also advertise Vision, Web Search, Web Reader, and Zread MCP for plan subscribers.[3][12]

Technical Architecture

Where code actually runs

The desktop client is the control plane. File edits, terminals, and Git run in the selected workspace: local disk, an SSH POSIX host, WSL on Windows, or a running Docker container. The first SSH connection installs a remote runtime under ~/.zcode/server. Later reconnects reuse it unless the version changes.[10]

This is remote execution of the same agent, not a cloud VM that returns a pull request. Model requests still leave the machine toward Z.ai, BigModel, or a configured third-party endpoint. The in-app HTTP proxy covers model APIs, MCP, WebFetch, and some subprocesses; it does not cover SSH connections or the Web Remote Control channel.[3][10]

Permissions are not a sandbox

Four modes: Ask before changes (default), Edit automatically, Plan, and Full access. Permission requests and plan approvals wait indefinitely. Ordinary clarifying questions auto-continue after five minutes unless that setting is off.[9]

Terms of Service are explicit about residual risk: models can misunderstand destructive filesystem instructions; the product asks for elevated local privileges; and ZCode recommends a virtual machine or sandbox rather than claiming to provide one. Third-party skills, plugins, and MCP servers are the user's responsibility.[8]

Isolation summary

BoundaryWhat the sources establish
OS / installDesktop builds for macOS (Apple Silicon and Intel), Windows (x64 and ARM64), and Linux (AppImage, deb, rpm). Current advertised version 3.14.3.[5][4]
Process isolationPermission UI plus local allow/deny rules. Not a documented hypervisor or OS sandbox.[9][8]
NetworkModel and MCP traffic to configured providers; optional HTTP proxy with documented exceptions.[3]
DataConversation inputs include code, files, logs, and tool results. September 2026 snapshot uploads are a separate, now-removed pipeline discussed below.[8][7]

Strengths

  • A first-party GLM harness, not only a plan that plugs into someone else's CLI. Execution modes, Goal Mode, browser, subagents, and AGENTS.md are productized around ZCode Agent.[1][2]
  • Real remote workspaces. SSH, WSL, and Docker keep the GUI on your laptop while the agent uses the machine that has the toolchain or GPU.[10]
  • Model flexibility after the GLM default. Official docs cover Z.ai, BigModel, Anthropic, OpenAI, OpenRouter, Moonshot, MiniMax, Xiaomi MiMo, and custom OpenAI/Anthropic-compatible providers.[3]
  • A free client and a documented trial before paying for a Coding Plan.[11][3]

Cautions

  • Local execution with model-grade judgment. Treat the agent as a privileged user. The vendor says so in the terms.[8]
  • Desktop installer versus source CLI. Do not assume every GitHub README surface is what the download page ships today.[5][6]
  • Regional billing and credit math. Global USD plans and China CNY BigModel plans are different checkouts. Credits, 5-hour windows, weekly caps, and peak-hour multipliers (Singapore time) change effective cost.[12][3][14]
  • Trust after the snapshot incident. Open-sourcing and stripping upload code are responses, not proof that earlier cloud objects were destroyed or that binaries match the public tree.[7][15]

What Developers Say

ferstar, September 18-23, 2026, published a first-hand reverse-engineering write-up after noticing ~/.zcode using hundreds of megabytes. On client 3.12.3, while logged in, the app packaged workspace snapshots (largely .git history) into .enc archives, requested OSS credentials from zcode.z.ai, and attempted Aliyun OSS uploads. UI toggles for "Optimize Experience" and "Repo Snapshot Indexing" did not stop packaging. A 313MB commercial-project archive failed size limits and never left the author's network; a small public repo snapshot was accepted. Z.ai's September 18 statement, quoted on Hacker News, attributed uploads to codebase indexing / Repo Wiki, said cloud copies were destroyed after Wiki generation, and promised open source plus a quota reset. ferstar's September 21-23 updates report that zai-org/ZCode is a two-then-three-commit dump with PRs historically locked, that checkpoint code in the public tree is local Git, and that client 3.14.3 no longer contains the repoSnapshot / upload-credential pipeline in unpacked app.asar or live logs. The author still recommends a filesystem lock on ~/.zcode/v2/checkpoints because binaries can update independently of the public repo. These are one engineer's captures and code reading, not an independent lab audit, and they do not reconstruct what happened to any object that did reach OSS before deletion claims.[7][15]

Hacker News discussion of that article (339 points, September 18, 2026) includes the translated vendor apology and user reactions that closed-source harnesses should be treated as untrusted. Commenters compared the episode to Grok Build's earlier repository-upload controversy. Those comments are contemporaneous opinion, not measurements of current 3.14.3 behavior.[15]

No ZCode workload was executed for this profile.

Pricing & Licensing

Checked September 23, 2026.

Client: free to download and use. Model inference is billed separately.[11]

Trial: new users get a short GLM trial after connecting Z.ai or BigModel. ZCode's welcome page lists 5 days, 3M GLM-5.3 tokens/day and 2M GLM-5-turbo tokens/day. The Connect Models page lists 3M GLM-5.3 and 5M GLM-5.3-Flash per day for the same 5-day window. Treat those figures as conflicting vendor documentation and read the in-app provider page for the live quota.[2][3]

Global GLM Coding Plan (credits, from the plan overview): starts at $18/month. Lite / Pro / Max 5-hour credits are 2,000 / 12,000 / 28,000; weekly credits 10,000 / 60,000 / 140,000. GLM-5.3 and GLM-5.3-Flash are the supported coding models, with peak-hour multipliers (weekdays 14:00-18:00 Singapore time) and off-peak discounts. The overview also says the same plan can be used from Claude Code, Cline, and OpenCode. Confirm current USD prices on the subscribe page at purchase; this review used the overview's "$18" floor rather than an unreadable marketing table.[12][14]

China BigModel in-app plan cards documented on the Connect Models page: individual Lite CN¥118, Pro CN¥538, Max CN¥1,078, with team plans from CN¥598. Those amounts come from official doc screenshots, not an independently fetched BigModel checkout form.[3]

Licensing: product terms are a hosted/client ToS (effective June 15, 2026), not an open-source grant for the downloaded binary. The September 21 public repository is Apache-2.0. Matching version tags do not prove bit-for-bit identity between that tree and the Electron build.[8][6][7]

Competitive Positioning

AlternativeReason to evaluate it alongside ZCode
Claude CodeAnother foundation-lab first-party harness, Claude-only, with a mature CLI. GLM Coding Plan can drive Claude Code; that is not ZCode Agent.
TemboA platform for running several coding agents in cloud or self-hosted environments with team review. Relevant when the decision is operating agents, not picking one GLM desktop ADE.[16]
Grok BuildAnother lab client that faced a repository-upload controversy and a later open-source response. Useful as a process comparison, not a feature matrix.

ZCode and Tembo

ZCode is a local/remote ADE whose default agent and GLM plan come from the same lab. Tembo is an orchestration layer around Claude Code, Codex, Cursor, Amp, OpenCode, Pi, and similar harnesses, with isolated cloud or licensed self-hosted environments, previews, and team workflows.[16]

They overlap only at the job-to-be-done ("get an agent to change a repo"). ZCode does not become a Tembo-style multi-agent control plane by adding SSH. Tembo's public site, checked September 23, 2026, does not list ZCode as a supported harness. Do not infer a native integration.

Disclosure: Ry Walker is Tembo's co-founder and CEO. Tembo competes in adjacent coding-agent infrastructure, not as a Z.ai foundation-model lab.

Ideal Customer Profile

Best fit: developers who want a GLM-5.3-native desktop agent with files, terminal, and browser in one window; teams already buying a GLM Coding Plan; people who can put the workspace on a machine they accept as the trust boundary.

Needs additional verification: organizations that require a documented isolation boundary, a stable enterprise DPA for historical snapshot data, bit-identical open-source builds, or a CLI-only install path matching the GitHub README. Those requirements are not established by the current installer docs and ToS.

Viability Assessment

Z.ai is shipping frequent desktop releases (3.10 through 3.14.3 in roughly four weeks), maintains bilingual docs, and now publishes an Apache-2.0 monorepo with 6.5k GitHub stars as of this review.[4][6] The Coding Plan is a live paid product with individual and team SKUs.[12]

The open question is trust after default-on workspace packaging. Removal of the upload pipeline and a public source drop are material. They do not by themselves prove destruction of earlier OSS objects or that every future binary will match the public tree.[7]

Assessment

Evaluate ZCode if you specifically want Z.ai's own agent and GLM-5.3 in a desktop ADE, and you are willing to treat the client as privileged local software. Use Plan or Ask-before-changes modes, keep secrets off the workspace, and verify the live client version against the snapshot-incident timeline. For team operation of multiple harnesses, or for execution you do not want on a developer laptop, look at Tembo or a cloud-task product instead of stretching ZCode into that role.

Sources

Share on XShare on LinkedIn