claude-code-decompiled

[](https://www.npmjs.com/package/@anthropic-ai/claude-code) [](docs/)

Independent research reports on Claude Code v2.1.88, focused on architecture, agent behavior, permission design, prompt assembly, MCP integration, and context management.

This repository is intentionally documentation-only. It collects original analysis and commentary about a publicly distributed software package. It is not presented as a runnable source release or supported CLI build.

Language: English | 中文 | 日本語 | 한국어 | Español

---

Claude Code is one of the most sophisticated production AI coding agents available today. These reports aim to make its architecture and behavior easier to study:

• Agent developers can study production patterns for tool orchestration, context management, and recovery logic.
• Security researchers can inspect telemetry, remote control mechanisms, and permission boundaries.
• AI researchers can examine prompt assembly, model routing, and agent loop design.
• Engineers can use the reports as references when building CLI-based agent systems.

These reports are based on technical analysis of a publicly distributed npm package and are presented here as research commentary.

---

This repository is designed to answer a fairly specific set of questions about Claude Code:

• How is Claude Code architected internally?
• How does the main agent loop actually work?
• How are tools registered, validated, permission-checked, and executed?
• How is the system prompt assembled from static and dynamic parts?
• What telemetry and privacy-relevant data paths exist?
• How does the permission model work in practice?
• How do