Build Your Claude Code From Scratch

Build Your Claude Code From Scratch is an educational codebase designed to teach the development of AI agents, progressing from basic tool-calling scripts to advanced multi-agent architectures with external protocol integration. Structured as the Agent Learning Project, it is organized into eight chapters, each introducing new components and building on the concepts of the previous ones. The project addresses the need for a step-by-step, hands-on approach to understanding and implementing conversational AI systems with increasing complexity.

The codebase covers a range of fundamental and advanced topics in agent development. Early chapters introduce tool-calling patterns, integration with JSON Schema and XML formats, and the ReAct (Reasoning and Acting) methodology, including environment configuration and user approval steps. Subsequent chapters add real-time streaming responses, dual-mode API clients, conversation history management with automatic compression and token tracking, and smart context cropping for prompt management and safety. Further progression includes structured task management, sub-agent delegation with isolated contexts and recursion, and integration with the Model Context Protocol (MCP) for external tool discovery and multi-server support.

The system architecture features several singleton managers that maintain consistent state and orchestrate the agent's operation. Key components include the Conversation manager (handling the ReAct loop, message recursion, tool execution, and streaming), APIClient (supporting OpenAI-compatible API calls and cost tracking), HistoryManager (managing conversation history and context windows), ToolManager (registering and proxying tool execution, including built-in and MCP tools), UIManager (handling user interaction and streaming output), PromptManager (supplying prompts for different agent contexts), and CostTracker (monitoring API usage and costs). The tool ecosystem employs a plugin architecture, allowing tools to be either built-in Python classes or dynamically loaded from external MCP servers. An example built-in tool is CmdRunner, which executes shell commands with timeout controls and user approval.

This project is intended for individuals seeking a comprehensive, practical guide to building conversational AI agents, from foundational concepts to advanced orchestration and protocol integration. Each chapter is self-contained and can be run independently, providing incremental learning and demonstration of increasingly sophisticated agent behaviors.