Z Code Review: AI Coding Environment by Z.AI With Agents and MCP Support

Z Code is a desktop application from Z.AI (the company behind GLM models) that bundles a file manager, terminal, Git panel, live browser preview, and AI agent chat into one interface. You open a project, tell the agent what you need, and it reads your files, writes code, runs commands, and commits changes.

I’ve been testing it over the past week and the thing that surprised me most is how different it feels from Cursor or Windsurf. Those tools are editors with AI bolted on. Z Code starts from the agent and builds everything around it. Z.AI calls it an Agentic Development Environment (ADE), which is a mouthful, but it’s accurate. The whole UI assumes you’re talking to an agent, not editing files by hand.

What Z Code actually does

Here’s what you get when you open it. You interact through natural language, and the agent handles file operations, terminal commands, and browser previews.

Installation

Z Code is available as a desktop app for macOS and Windows. There’s no CLI-only version, it’s a full GUI application.

macOS

1. Download the DMG from zcode-ai.com
2. Open the DMG file
3. Drag Z Code.app to your Applications folder
4. Launch from Launchpad

Windows

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

After installation, open the app and log in to start using it.

Download Z Code

Configuring AI providers

The provider setup is one of the better parts of Z Code. You’re not locked into a single AI service. You can connect multiple providers and switch between models per conversation.

Supported providers

Provider	Models Available	How to Connect
Z.AI / BigModel	GLM-5, GLM-4.7, GLM-4.6, GLM-4.5	Built-in, bind your Zhipu account
Anthropic	Claude Sonnet 4.5, Opus 4.5, etc.	Enter Anthropic API key
OpenRouter	100+ models	Set base URL to https://openrouter.ai/api, add API key
Moonshot	Kimi K2.5, Kimi K2-Turbo	Base URL: https://api.moonshot.cn/anthropic
Custom	Any OpenAI/Anthropic-compatible API	Add provider name, base URL, and API key

Configuration steps

There are three ways to add a provider:

1. Via Connect button: Click “Connect” on the welcome page to bind your Zhipu account or set a third-party API key.
2. Via Model Selector: Click the model name in the chat box, then “Manage Models” at the bottom.
3. Via Settings icon: Click the settings gear in the top-right corner to go directly to the Models panel.

Permission modes

Every time you start a chat, you pick how much freedom the agent gets. Z Code has four permission modes, and you can cycle through them with Shift + Tab.

Mode	What Happens	Good For
Always Ask	Agent asks before every action (file edits, commands, network)	Critical tasks needing full control
Accept Edits	Agent edits files freely but asks before running commands or accessing the network	Trusted code generation, controlled system access
Plan Mode	Agent creates an action plan first, waits for your approval before executing	Complex tasks where you want to review the approach
Bypass Permissions	Agent does everything automatically with no confirmation	Isolated environments only, higher risk

Specialized agents

Z Code ships with built-in agents that the main agent can hand work off to. Each runs in its own context window with independent prompts and tool permissions.

Built-in agents

Agent	What It Does
bug-analyzer	Deep code execution flow analysis and root cause investigation
code-reviewer	Security vulnerability detection, performance issues, production reliability checks
dev-planner	Breaks requirements into tasks with dependency analysis
story-generator	Creates user stories with acceptance criteria from git diffs or PRDs
ui-sketcher	Transforms feature requirements into ASCII interface designs

Invoking agents

Type @ in the chat to see the agent list, or reference one directly:

@code-reviewer Review the payment module
@bug-analyzer Investigate this error
@dev-planner Break down the user registration feature

Agents can also run in parallel. Ask the main agent to “review API changes and update documentation” and it’ll spin up multiple agents simultaneously.

Creating custom agents

You can define your own agents via the Settings UI or as Markdown files:

Project-level agents go in .zcode/agents/ (shared with the team via Git). User-level agents go in ~/.zcode/agents/ (available across all your projects).

Here’s an example agent file:

---
name: security-auditor
description: Security expert. Use when implementing authentication, payments, or handling sensitive data.
model: inherit
---

You are a security expert responsible for reviewing code vulnerabilities.

When executing:
1. Identify security-sensitive code paths
2. Check for common vulnerabilities (injection, XSS, authentication bypass)
3. Verify secrets are not hardcoded
4. Review input validation and sanitization

Report findings by severity:
- Critical (must fix before deployment)
- High (fix as soon as possible)
- Medium (address when available)

Skills system

Skills are Markdown files that teach the agent how to handle specific tasks. They work like reusable playbooks. When your request matches a skill’s description, the agent either suggests it or you pick it with @.

Skill types

Type	Storage	Scope
User Skills	~/.claude/skills/	All projects
Plugin Skills	Installed via plugins	Auto-available
Project Skills	.claude/skills/	Current project only

Creating a skill

A skill is a Markdown file with YAML frontmatter:

---
name: explaining-code
description: Explain code with visual diagrams and analogies
---

# Code Explanation Standards

When explaining code, always include:

1. **Start with analogies**: Compare code to everyday things
2. **Visual diagrams**: Use ASCII diagrams to show flow
3. **Step-by-step breakdown**: Clearly explain each part

Skills can also contain executable scripts in Python, Node.js, or other languages. The agent runs the script and only the output consumes context tokens, not the script source.

MCP services

Z Code uses the Model Context Protocol to extend what agents can do. Three built-in MCP services cover common needs:

Plugin system

Plugins add new commands, agents, MCP connections, language servers, and more to Z Code. It ships with access to the Claude Code official plugin marketplace.

Plugin types

Type	What It Adds
Agent	Specialized agents with domain workflows
Command	Custom / slash commands
MCP	External tool and data source connections
LSP	Language servers for code navigation and type checking
Skill	Behavior guidelines for the agent
Hook	Auto-execute actions on specific events

Installation scopes

Scope	Location	Visibility
User	~/.claude/	All your projects
Project	.claude/settings.json	Shared with team (commit to repo)
Local	.claude/local.settings.json	You only, current project

You can also add third-party or private team marketplaces by pointing Z Code at a Git repository or a hosted marketplace.json file.

Output styles

Z Code lets you pick how the agent talks to you. Five built-in styles ship with it:

Style	Best For	Description
Default	Daily work	Concise and professional
Explanatory	Understanding code	Detailed explanations of implementation choices
Learning	Learning new tech	Guided teaching with hands-on practice pauses
Coding Vibes	Casual sessions	Energetic tone with emojis and slang
Structural Thinking	Architecture design	Layered, modular analysis

You can create custom styles too. Each style is a Markdown file with frontmatter and body instructions that define the agent’s tone, format, and behavior.

Version control

Z Code handles versioning at two levels, and honestly this is one of the features I liked most.

Conversation-level rollback

Every message you send creates a checkpoint. You can:

• Review: See all file changes after any interaction in a multi-file diff view
• Undo: Revert only the changes from the last interaction
• Restore: Jump back to the state after any specific message in your chat history

This means your conversation history is also your version timeline. Try different approaches, and roll back if something doesn’t work.

Built-in Git panel

A full Git management panel is built into the left sidebar:

• View modified files with status markers (M for modified, U for untracked)
• Write commit messages and commit with one click
• Switch branches from the bottom status bar
• Browse commit history

Memory system

Z Code reads a CLAUDE.md file from your project root and loads it into every conversation. Put your project conventions, architecture notes, and coding standards here.

Quick start

Run /init in the chat and Z Code will generate a CLAUDE.md file for your project automatically.

What to put in CLAUDE.md

# Project Architecture

This project uses a layered architecture:

- `src/api/`: API route definitions
- `src/services/`: Business logic layer
- `src/models/`: Data models

## Development Conventions

- API routes use RESTful style
- All database operations go through the services layer
- Error handling uses the unified errorHandler utility

## Code Style

- Variables use camelCase
- Constants use UPPER_CASE
- File names use kebab-case
- 2-space indentation
- Single quotes for strings

Keep it under 200 lines. Put the most important information first. Don’t include secrets or API keys.

Remote development

Z Code can stream your desktop session to a phone over WebSocket. It’s useful for checking on long-running tasks or handling a quick fix when you’re not at your desk.

1. Click the “Remote Development” icon in Z Code’s top-right corner
2. A QR code appears on screen
3. Scan it with your phone
4. Your phone shows a live view of the Z Code session

All code execution stays on the desktop. The phone only handles display and text input. Currently limited to one mobile device per session.

Z Code vs other AI coding tools

I put together a comparison table since this is probably the first question most people have:

Feature	Z Code	Cursor	Claude Code CLI	Windsurf
Type	Standalone ADE	IDE (VS Code fork)	Terminal CLI	IDE (VS Code fork)
Multi-Model	Yes, GLM, Claude, GPT, Kimi, DeepSeek, custom	Yes, multiple providers	Claude only	Yes, multiple providers
Specialized Agents	Built-in + custom	No	No	AI Flows
Skills System	Yes, reusable Markdown playbooks	No	No	No
Plugin Marketplace	Yes, agents, commands, MCP, LSP	Extensions	No	Extensions
Chat Versioning	Yes, conversation checkpoints + rollback	No	No	No
Remote Dev (Mobile)	Yes, QR code to phone	No	No	No
MCP Support	Yes, built-in + custom	Community	Yes	Yes
Memory Files	CLAUDE.md auto-loaded	.cursorrules	CLAUDE.md	.windsurfrules
Git Integration	Built-in panel	Built-in	CLI-based	Built-in

Getting started in 5 minutes

The quickest way to get going:

Step 1: Download and install Z Code from zcode-ai.com

Step 2: Open the app and click “Connect” to set up your AI provider. If you have an OpenRouter key, that gives you access to the widest range of models.

Step 3: Open a project folder in Z Code.

Step 4: Type /init in the chat to generate a CLAUDE.md memory file for your project.

Step 5: Start chatting. Ask the agent to explain your codebase, fix a bug, or build a new feature.

Get Started with Z Code

Pros and cons

Drawbacks

• Desktop only: No web version, no VS Code extension. You use their app or nothing.
• New tool: The ecosystem is still small. Fewer community resources compared to Cursor or Copilot.
• Z.AI ecosystem: Some MCP services need Zhipu/BigModel API keys specifically.
• macOS and Windows only: No Linux support yet.
• Lots of moving parts: Agents, skills, plugins, MCP, memory. There’s a learning curve to figuring out which piece to use when.

Frequently asked questions

Do I need Claude Code to use Z Code?

No. Z Code supports multiple AI providers. You can use GLM-5, models through OpenRouter, the Anthropic API directly, Moonshot (Kimi), DeepSeek, or any service that speaks the OpenAI or Anthropic protocol. Claude Code is just one optional connection method.

Is Z Code free?

Z Code itself is free to download and use. You pay for the AI models you connect — either through Z.AI’s GLM Coding Plans, OpenRouter credits, Anthropic API usage, or whichever provider you choose.

Can I use Z Code with local models?

Yes. Z Code supports any OpenAI or Anthropic-compatible API endpoint. Point it at a local inference server (like Ollama or vLLM) and it’ll work with whatever model you’re running.

How does Z Code compare to Cursor?

Cursor is a VS Code fork focused on AI-assisted editing. Z Code is a standalone environment built around agent interactions. Z Code has specialized agents, a skills system, a plugin marketplace, and conversation-level version rollback. Cursor has a more familiar IDE experience and a larger existing user base.

Does Z Code work on Linux?

Not at the time of writing. Z Code is available for macOS and Windows only.

Wrapping up

Z Code takes a bet that the IDE of the future is agent-first, not editor-first. If that bet pays off, it has a head start over tools that are retrofitting agents into existing editors. The multi-model support, custom agents, and conversation versioning are genuinely useful, and I found myself reaching for the rollback feature more than I expected.

The trade-off is that you’re leaving behind the VS Code ecosystem and its extensions. For some projects that won’t matter. For others it’s a dealbreaker. My suggestion: try it on a side project first. The free download and BYOK pricing model make it low-risk to test.

Try Z Code