ZCode AI Review 2026: Free AI Coding Agent vs Cursor

Disclosure: this guide contains referral links. If you buy through them, I may earn a commission at no extra cost to you.

In this updated ZCode review I look at Z.AI’s free AI coding agent as it stands in November 2026: what changed in v3.17.2, how Goal Mode handles long unattended tasks, and where the whole package lands against Cursor. If you want a free desktop coding agent that keeps working while you walk away, ZCode is the one I keep coming back to: hook up a model, set a goal, and let it iterate.

ZCode is the official desktop app from Z.AI, the team behind GLM. File manager, terminal, Git, browser preview, and agent chat sit in one window. Open a project, describe what you need, and the agent reads files, writes code, runs commands, and commits. The harness went open source under Apache-2.0 at github.com/zai-org/ZCode. It had 8,638 stars and 2,695 forks as of mid-November 2026. The app itself is at v3.17.2 (released November 3, 2026).

Since my last update of this review: a Sites plugin for one-click website deploys, workflow auto-resume after quota resets, MCP servers that start without Node, a batch of plan-mode security hardening, and a macOS 13+ floor. The full changelog is in the next section. For the wider field of agents and IDEs, see my best AI coding tools and agents compared roundup.

What shipped since late October (v3.17.1 to v3.17.2)

ZCode has been shipping multiple releases a month. If you tried an older build earlier this year, here is the short version of where things stand. The harness was open-sourced and the legacy snapshot/upload pipeline was stripped out back in v3.14.0. Both good moves that raised the trust floor. Since then the pace continued. Between v3.17.1 (October 30) and v3.17.2 (November 3-4, 2026):

What is ZCode? Z.AI’s free, open-source AI coding agent

ZCode calls itself an “Agentic Development Environment”. Marketing language, but it matches how the app behaves. The agent is the product; the UI follows. Cursor and Windsurf are editors with chat stapled on. ZCode inverts that: you open a workspace, describe a task, and the agent drives file operations, terminal commands, Git, and browser previews while you watch (or don’t).

Here’s the workspace you see when you open it:

For model context (what GLM-5.3 is competing against and why a 1M-context open-weights model matters), see my best open-source LLMs vs Claude Opus comparison.

The ZCode Agent and AGENTS.md

The agent handles complex project understanding, long-task planning, and continuous code changes across multiple files. GLM-5.3’s stable 1M context window means the agent can hold an entire mid-size codebase in view without constant re-reads.

To add context, click the + button at the bottom-left of the chat input:

• Attachments upload screenshots, documents, and requirement material
• @ mentions reference files in the workspace so the agent can locate relevant code
• # conversations link a past conversation to bring its context into the current task
• / commands call saved prompts to reuse a fixed workflow
• $$ skills invoke reusable skill playbooks

For project instructions, ZCode reads an AGENTS.md file from your project root and loads it into every conversation. Put your project conventions, architecture notes, and coding standards here.

Execution modes

Five modes control how much freedom the agent gets. Press Shift + Tab while the chat input is focused to cycle through them.

From v3.17.2, plan mode got a security pass: high-risk operations can no longer be allowed even in plan mode, and subtask permissions are restricted to match the main session. If you were relying on plan mode as a safety net for dangerous commands, it now acts like one.

Skills system

Skills are Markdown files that teach the agent how to handle specific tasks. They work like reusable playbooks.

MCP services

ZCode uses the Model Context Protocol to extend what agents can do. The exclusive MCP set is now four services: Visual Understanding, Web Search, Web Reader, and the newer Zread. From v3.17.2, MCP servers start without Node installed (bundled runtime), and there is a draft MCP configuration GUI if you prefer not to hand-edit JSON.

Plugin system

Plugins add new commands, agents, MCP connections, language servers, and more. ZCode ships with access to the Claude Code plugin marketplace, so you can install community plugins you’re already familiar with. From v3.17.2 there’s a one-click update all plugins button, and the document viewer is now isolated per session (less cross-talk between parallel tasks).

The new Sites plugin (v3.17.2) is worth calling out: one-click website deployment from inside ZCode, with no separate CI pipeline needed for simple static deploys.

Version control

ZCode handles versioning at two levels.

At conversation level, every message creates a checkpoint. Review all file changes after any interaction in a multi-file diff view, undo only the last interaction’s changes, or jump back to the state after any specific message.

The built-in Git panel shows modified files with status markers. You can write commit messages and commit with one click, switch branches, and browse commit history.

Installation

ZCode is a desktop app for macOS (Apple Silicon and Intel), Windows (x64 and ARM64), and Linux (x64 and ARM64).

macOS

1. Download the DMG from zcode.z.ai
2. Open the DMG file
3. Drag ZCode.app to your Applications folder
4. Launch from Launchpad

macOS 12 (Monterey) and earlier will not run v3.17.2+. The app needs macOS 13 Ventura minimum. If you see a “you cannot open this app” error on an older OS, that’s the cause.

Windows

1. Download the installer from zcode.z.ai
2. Double-click the installer
3. Follow the setup wizard

v3.17.2 fixed Win64 installer compatibility and added taskbar/tray unread notifications.

Linux (DEB, RPM, AppImage)

Linux is fully supported. Pick between AppImage, DEB, and RPM builds for x64 and ARM64 on the download page. Closing the window keeps the app in the system tray, which is what lets scheduled and idle-time tasks keep running in the background.

Linux troubleshooting

Common failure modes on Linux and WSL:

AppImage won’t start: dlopen(): error loading libfuse.so.2. New distros ship FUSE 3 only. Fix:

sudo apt install libfuse2

Keep the AppImage at a fixed path (e.g. ~/Applications/ZCode.AppImage) so desktop integration does not break after updates.

NVIDIA/GPU startup errors. Launch with software rendering if the GPU driver stack is fighting Electron.

WSLg quirks. Launch ZCode and the browser from the same WSL environment. CJK input needs IBus or Fcitx5 configured inside WSL.

Verify your download and first launch

GitHub releases attach a SHA256SUMS file. Verify before you run anything:

sha256sum -c SHA256SUMS

Then first launch: the app opens, the workspace loads, and you see the welcome screen with the file tree and chat panel. If the window opens but the workspace is blank, restart once. First-run indexing can lag on very large repos.

Configuring AI providers

ZCode supports multiple providers. You are not locked into a single service.

Supported providers

If you want to run GLM models in a different harness, you can also run GLM 5.2 and other models in the Codex app. The endpoint setup is similar.

Endpoint reference for custom setups:

Terminal environment variables do not sync into ZCode. Configure providers in-app (Quick Connect or manual Base URL + key).

Verify: check your plan binding (Usage Stats)

GLM Coding Plan pricing (2026)

Z.AI’s GLM Coding Plan was relaunched in 2026 with a credit-based system. Plans start at $18/month. Here are the hard numbers as of November 2026:

Credit math: model credits = (input tokens x in-multiplier + cached-input tokens x cached-multiplier + output tokens x out-multiplier) / 10,000. MCP tool credits = calls x out-multiplier with no /10,000 division, so one Web Search, Web Reader, or Zread call costs 1.2 credits.

The peak window is Monday to Friday 14:00-18:00 SGT (UTC+8), charging the standard 1x credit rate. Everything else (nights, weekends, and the remaining weekday hours) charges 0.5x. For readers in Europe/Bucharest, peak is 08:00-12:00 in winter (EET) and 09:00-13:00 in summer (EEST). Scheduling your heavy agent sessions outside that window effectively doubles your plan size for the same money. This is the single biggest lever on plan cost.

The estimated weekly token allowance at 95% cache hit (GLM-5.3) is Lite 48-97M, Pro 290-580M, Max 676-1,352M tokens per week (range = off-peak max vs peak min). Z.AI claims “up to 92% savings vs pay-as-you-go GLM-5.3 API”. The API prices are $1.40 input / $4.40 output per 1M tokens ($0.26 cached input), so the math checks out for sustained use.

For model routing, all plans support GLM-5.3 + GLM-5.3-Flash. Requests for GLM-5.2 or GLM-5.1 auto-route to GLM-5.3; GLM-4.7 requests auto-route to GLM-5.3-Flash. You always get the current-generation model.

When quota runs out, plan tools simply stop until the 5-hour cycle refreshes. Quota exhaustion never touches your account balance. There is no automatic fallback to pay-as-you-go. Quota Reset Cards (5-hour and weekly variants) exist as campaign perks.

Billing details worth knowing: subscriptions are non-refundable. Cancel at least 24 hours before renewal. Same-tier upgrades stack (e.g. Lite monthly to Lite annual = 13 months total); cross-tier upgrades take effect immediately with pro-rated conversion to account balance. Off-peak plans no longer auto-renew after expiration (v3.17.2).

If you’re looking at this tier of pricing versus GitHub Copilot’s usage-based model, I break down the cheaper alternatives to GitHub Copilot elsewhere on the site. And for another option in this space, Freebuff, another free AI coding agent we tested offers 30h/day free with a different tradeoff profile.

Goal Mode: set an objective and walk away

This is the feature that keeps me on ZCode. Type /goal <objective> and the agent keeps iterating toward a verifiable goal. You don’t type “continue” every few minutes or babysit the run.

The /goal command cheat sheet:

/goal                    Show the current goal and status
/goal <objective>        Set or replace the session goal
/goal replace <obj>      Explicitly replace the current goal
/goal pause              Pause goal-driven iteration (keeps state)
/goal resume             Resume from where you paused
/goal clear              Remove the goal entirely

One goal per session. Plain-language goal changes work too; you don’t need the exact subcommand syntax.

Verification is evidence-based. A plan, a checklist, or elapsed effort does not count as “done.” The agent looks at changed files, command output, and test results. Any unfinished to-do item blocks the verdict from being “complete.” Checklist items stay in the iteration where they first appeared, so you get a readable history of what was tried and when.

State persists across interruptions. Pausing keeps completed rounds, tool history, and files. Closing the session and reopening restores the goal. Stopping a running task auto-pauses the goal so it doesn’t burn quota unattended. Combined with v3.17.2’s no-progress auto-stop, Goal Mode is now safe to leave running overnight. If the agent gets stuck in a loop, it stops and notifies instead of draining credits.

Good use cases:

• “Refactor the whole module and keep the tests passing”
• “Fix all TypeScript compile errors”
• “Get this page’s Lighthouse performance score above 90”

The summary panel in the top-right shows goal status, elapsed time, total tokens, and iteration count. Pair Goal Mode with Full Access or Auto Edit to minimize interruptions and keep iterations flowing.

Automations, Browser Automation & Repo Wiki

Three features that don’t always make the highlight reel but earn their keep in daily use.

Automations (scheduled and idle-time tasks)

Automations come in two flavors. Scheduled tasks take a time and a repeat rule; results land in a session. Create one via a form, “Create in chat,” or from four built-in templates: weekday standup digest, daily risk scan, Friday release brief, Wednesday doc sync check. Idle-time tasks have no schedule. The agent runs them on spare capacity for free, without eating plan quota (rolling out to subscribers).

You can also just describe the schedule in plain language: “Every weekday at 9am, summarize this project’s code changes and open follow-ups.” On Linux, remember the app needs to stay in the system tray for background tasks to run.

Browser Automation (Browser Use plugin)

The agent can drive the built-in browser panel: open URLs, click, fill forms, scroll, take screenshots, and react to page state. The official Browser Use plugin is on by default and ships two skills: one for general browser driving, one for walkthrough-style UI testing. Desktop-only for now. This is a step beyond the passive browser preview the app had at launch. The agent can interact with pages.

Repo Wiki

Point ZCode at a codebase and it generates an architecture wiki per workspace. Every claim carries a source location you can click through to. That helps you verify the agent understood the code before you trust its refactoring plan. Language and model are selectable. Access it via the file-tree icon. Great for onboarding to an unfamiliar repo.

Bot Channel & Remote Control

Two ways to keep tabs on ZCode when you’re away from your desk.

Bot Channel connects WeChat and Feishu to your ZCode workspace. Message the bot from your phone’s chat app to check task progress, send instructions, and keep the agent moving.

Setting up Feishu:

1. Open Bots settings in ZCode’s left sidebar
2. Click Create Bot and select Feishu
3. Scan the QR code. ZCode creates the Feishu app automatically
4. Send /bind pairing-code in the Feishu conversation
5. The bot is ready to use

WeChat works similarly. Scan to sign in and bind automatically. Since v3.15 the bot can collaborate inside Feishu group chats and standalone topics. v3.17.2 fixed result-push notifications for WeChat automation tasks. Z.AI has said DingTalk, Discord, and WeCom support is coming, but none of those are shipped as of v3.17.2. Don’t buy based on the roadmap alone.

Remote Control lets your phone connect to the current desktop workspace. Scan a QR code, and your phone shows a live view of the session. Read agent responses, send new instructions, confirm or stop long-running tasks.

The desktop stays as the runtime. The phone only handles display and text input. One phone connection per session. v3.17.2 added account login for mobile Remote Control, a device manager in settings, a compact searchable remote-device list, and the ability to stop/restart remote services from the menu. For longer-lived access from chat apps, use Bot Channel instead. If you want a similar phone-to-agent setup with different tooling, I covered how to control AI agents from your phone with Moshi + Herdr in a separate guide.

ZCode vs Cursor and other AI coding tools

The honest positioning: Cursor gives you a familiar IDE experience, a larger extension ecosystem, and a bigger community. ZCode gives you agent-first workflows: Goal Mode, multi-agent orchestration, Bot Channel, conversation-level rollback, and an open-source harness. If your workflow is “edit code in a VS Code fork with AI autocomplete,” Cursor wins. If your workflow is “describe a task, walk away, come back to a result,” ZCode wins.

For the full picture including other tools (Claude Code, Windsurf, Amp, and more), my best AI coding tools and agents compared covers the field. If you were using Amp Code and got cut off by the free tier pause, Amp Code Free is paused: what to use instead explains the situation and why ZCode is a natural replacement.

If you want a single API key to route across Claude Code, OpenAI Codex, and Gemini CLI, Agent Router does the multi-model plumbing, useful if you want the model flexibility of ZCode but prefer a terminal-first harness.

Pros and cons

What I like:

What I don’t:

• The plugin ecosystem is still smaller than Cursor’s extension marketplace, though UI plugins (MCP Apps) and the Sites plugin are closing the gap
• Some MCP services need Zhipu/BigModel API keys specifically
• The learning curve is real: agents, skills, plugins, MCP, Goal Mode, workflows, execution modes. There’s a lot to figure out.
• Bot Channel only supports WeChat and Feishu right now (Discord and others are promised but unshipped)
• macOS 13+ requirement from v3.17.2 leaves older Macs behind
• GLM-5-turbo (the trial’s second model) is murky. It only appears in the trial grant and is not on the public API price list

Frequently asked questions

Wrapping up

This ZCode review covers the app as of v3.17.2 (November 2026): a free, open-source AI coding agent with Goal Mode that genuinely handles unattended work. The harness is Apache-2.0, Linux support is solid, and the GLM Coding Plan starts at $18/month with off-peak credits at half price.

Who it’s for: GLM users, anyone running long unattended tasks (refactors, test-fix loops, Lighthouse targets), and people who want to check on agent progress from their phone. Who should stay on Cursor: developers who live in an IDE-native workflow and want the largest extension ecosystem.

The cost picture is straightforward: free harness, model costs from $18/month Lite (or bring your own API key). Scheduling heavy sessions outside the Mon-Fri 14:00-18:00 SGT peak window doubles your effective plan size. The 5-day free trial is enough to see if Goal Mode fits how you work.