MoAI-ADK Documentation

The moai CLI (the Go binary) is organized into three functional groups:

1. Launch: Commands to start Claude Code sessions with specific backends (moai cc, moai cg, moai glm).
2. Project: Commands for lifecycle management, including initialization, updates, and diagnostics (moai init, moai update, moai doctor, moai status).
3. Tools: Advanced utilities for configuration, inventory, hooks, worktrees, and specifications (moai profile, moai inventory, moai hook, moai worktree, moai spec, moai harness).

Use moai version to check your current installation.

moai version

The moai CLI (a Go binary) is a tool for managing agentic AI development workflows. It is distinct from the /moai slash command used within Claude Code's interactive interface. The CLI is organized into three main command groups:

1. Launch: Commands to start Claude Code sessions with specific backends (moai cc, moai cg, moai glm).
2. Project: Commands for project lifecycle management, including initialization, updates, diagnostics, and status checks (moai init, moai update, moai doctor, moai status).
3. Tools: Utilities for configuration, inventory, hooks, worktrees, and more (moai profile, moai inventory, moai hook, etc.).

Use moai version to check your currently installed version.

Claude Code is a model that reasons about code with built-in tools for file editing, search, and execution, supplemented by layers for context, extension, and automation.

Its features are organized into four functional groups:

1. Foundations: Core capabilities like code editing, search, command execution, and slash commands.
2. Context and Memory: Managing the session state via CLAUDE.md and the context window.
3. Extensibility: Expanding capabilities using Skills, MCP (Model Context Protocol), Hooks, and Plugins.
4. Agents and Automation: Scaling work via Subagents, Agent teams, Worktrees, and Checkpoints.

MoAI-ADK provides orchestration primitives for moving beyond single-conversation interactions toward autonomous agentic loops. Developers can delegate tasks to multiple workers, facilitate collaboration within teams, and manage large-scale work via scripts.

Core orchestration primitives include:

• Subagents: Delegated workers operating in isolated contexts.
• Agent Teams: Collaborative groups typically consisting of 3-5 members.
• Dynamic Workflows: Script-based, large-scale orchestration for fanning out work.

To build reliable autonomous execution loops, these primitives are combined with Worktrees (for isolation), Goal-Directed Execution (using /goal to run until a condition is met), Scheduled Tasks (for recurring background runs), and specific strategies for navigating Large Codebases.

MoAI-ADK (Agentic Development Kit) is a strategic orchestration framework designed for Claude Code. It focuses on three primary axes to move from 'vibe coding' to 'harness engineering':

1. Tokenomics: Focuses on reducing inference costs (by 60-70%) using techniques like context dieting and prompt caching.
2. Agentic Loop Engineering: Enables autonomous improvement cycles where agent loops accumulate observations to evolve harness guidance through recursive self-learning.
3. Agentic Harness: Provides a composable execution environment utilizing skills, hooks, and MCP (Model Context Protocol) for extensible orchestration.

Key features include the MoAI Orchestrator for task delegation, SPEC-Based TDD/DDD methodologies, the TRUST 5 Framework (Tested, Readable, Unified, Secured, Trackable), and Progressive Disclosure for efficient skill loading.

MoAI-ADK (Agentic Development Kit) is an agentic harness designed for Claude Code. It acts as a system that wraps around the model to enforce constraints that the model itself cannot maintain due to its probabilistic, token-based nature.

While Claude Code handles the generation, MoAI-ADK provides the structural layer for:

• Cost (Tokenomics): Optimizing token usage and managing model routing to ensure high quality at predictable costs.
• Self-Improvement (Agentic Loop Engineering): Turning observations from previous sessions into rules to prevent repeating mistakes.
• Quality Control (Agentic Harness): Using lifecycle management (SPEC) and quality gates (TRUST 5) to prevent expensive rework.

It is distributed as a single Go binary that runs on macOS, Linux, and Windows without external dependencies.

MoAI provides three primary launchers to run Claude Code with different backend configurations. Choosing a launcher is a tokenomics decision based on your cost and reasoning requirements.

Note on GLM limitations:

• GLM does not support the auto permission mode.
• Z.AI has low concurrent-request limits (1-3 in-flight on paid tier). For parallel multi-agent execution, moai cg is more stable.

A Claude Code plugin is an extension bundle that packages multiple elements—such as commands, agents, skills, hooks, and MCP—into a single versioned directory for sharing and reuse.

Unlike standalone configuration in the .claude/ directory, plugins use a manifest file (.claude-plugin/plugin.json) to provide an identity and automatic namespacing. For example, a skill named hello in a plugin named my-plugin is invoked as /my-plugin:hello to prevent collisions with other configurations.

{
  "name": "my-first-plugin",
  "description": "A greeting plugin to learn the basics",
  "version": "1.0.0",
  "author": { "name": "Your Name" }
}