Claude Code
Internals
Everything about how Claude Code works, extracted directly from the source code and explained for everyday users.
Explore all sections
System Prompt
Modular prompt hierarchy and key rules
Tools
45 built-in tools by category
Agents
6 built-in agent types and capabilities
Memory
Persistent memory across sessions
Permissions
6 modes and the YOLO 2-stage classifier
Context & Compaction
Token budget, auto-compact, output limits
Costs
Pricing per model and the /fast mode trap
Retry Logic
Exponential backoff and error handling
Proactive Mode
KAIROS autonomous mode and tick prompts
Hidden Features
Feature flags and undocumented capabilities
Practical Tips
Top 10 tips derived from the source code
Skills
17 bundled skills, custom skill format, and discovery system
Slash Commands
~100 slash commands across 5 categories
MCP
Model Context Protocol: transports, auth, lifecycle, and enterprise policies
Bridge & IDE
IDE integration (VS Code, JetBrains) and remote bridge to claude.ai
Coordinator Mode
Multi-agent orchestration: hub-and-spoke coordinator and peer-to-peer team swarms
YOLO Classifier
AI security classifier: 2-stage XML protocol, fast paths, fail-closed error handling, and 28 block rule categories
Hooks
27 events, 4 hook types (command/prompt/agent/HTTP), matching, permission integration, and async patterns
Plugins
9 capability types, plugin manifest format, marketplace sources, installation flow, and enterprise policies
Custom Agents
Markdown-based agent definitions: frontmatter schema, tool config, memory scopes, MCP integration, and 6 built-in agents
Bash Security
5 defense-in-depth layers, 22+ validators, permission pipeline, sandbox (seatbelt/bubblewrap), path validation, env scrubbing, and 12 security principles