Mini-Code Documentation

What is Mini-Code?

Mini-Code is an educational AI coding assistant that demonstrates how modern AI assistants work. It's designed to help you understand the architecture and workflows behind AI-powered coding tools.

Learning Objectives

• Understand AI Assistant Architecture: See how agents coordinate between LLMs and tools
• Learn Permission Systems: Explore multi-layer security models for AI tools
• Master Tool Integration: Understand how tools are executed with proper validation
• Study Async Workflows: See how permission requests and tool execution work together

Quick Navigation

Topic	Description	Learning Focus
System Architecture	High-level system design and workflow	Architecture
Configuration	Settings and API management	Customization
Tools	Tool system concepts and execution	Extensibility
Permission	Multi-layer permission system	Security
LLM Tool Integration	Complete integration flow	Coordination
User Interface	UI components, state management, and interactions	User Experience
Configuration	Settings, API management, and MCP integration	Customization

Key Learning Concepts

Mini-Code demonstrates several important concepts for building AI assistants:

1. Agent Coordination Pattern

The Agent Executor coordinates between LLM decisions and tool execution, creating a loop that continues until the task is complete.

2. Tool System Design

All tools implement a unified Tool<Input, Output> interface with intelligent execution:

• Read-only tools execute concurrently for better performance
• Writing tools execute sequentially to maintain consistency
• Permission integration ensures security before execution

3. Multi-Layer Permission System

Three-layer security model balancing safety with convenience:

• Approval Modes: Configurable auto-approval levels
• Project Policy: Persistent permissions stored in config
• Session Policy: Temporary grants for current session

4. Streaming Architecture

• Real-time responses: Users see LLM responses as they're generated
• Progressive tool rendering: Tools show results as they complete
• Async permission handling: Permission requests don't block the UI

5. MCP Integration

• Multi-transport Support: Connect to MCP servers via stdio, HTTP
• Tool Bridging: Automatically convert MCP tools to mini-kode's tool system

6. User Control Features

• Slash Commands: Direct commands for session management (/clear, /compact, /init)
• Auto-compaction: Automatic conversation compression when token limits approach
• Project Memory: Persistent project context via AGENTS.md file