How Claude Code Works Internally | Explained using claw-code | fastest repo to surpass 100K stars

About this lesson

Ever wondered what happens under the hood when you use Claude Code? In this deep dive, we reverse-engineer the complete architecture of Anthropic's agentic coding assistant — from the CLI entry point to the ConversationRuntime's agentic loop, the tool execution system, permission gates, pre/post hooks, API streaming, session compaction, MCP integration, and more. Claude Code was originally built in TypeScript and briefly exposed its internals to the world. Thanks to the open-source claw-code clean-room rewrite (in both Rust and Python), we can now trace every single component of this system. What you'll learn: • The full architecture map — every component and how they connect • How the CLI wires config, tools, prompts, and permissions together • The layered configuration system with user/project/local overrides • Dynamic system prompt construction with 8 distinct sections • ConversationRuntime — the heart of the agentic loop • run_turn() — the core loop that makes Claude Code autonomous • Built-in tool ecosystem: Bash, FileRead, FileWrite, Grep, Agent, and more • Permission system: ReadOnly → WorkspaceWrite → DangerFullAccess • Pre/Post tool hooks for custom safety guardrails • API streaming events: TextDelta, ToolUse, Usage, MessageStop • Session management and automatic compaction • MCP (Model Context Protocol) for connecting external tools • Slash commands and the Skills system • Complete end-to-end trace of a single prompt through every component This is the most comprehensive breakdown of Claude Code's architecture available anywhere. #ClaudeCode #AIEngineering #Anthropic #AgenticAI #CodingAssistant #LLM #MachineLearning #SoftwareArchitecture #AITools #DeveloperTools #ClaudeAI #RustProgramming #SystemDesign #AIDeepDive #TechExplained

Original Description