claude-code-internals

TypeScript source code extracted from the Claude Code CLI (@anthropic-ai/claude-code). Useful for understanding how Claude Code works under the hood.

What's in here

The src/ directory contains the full CLI source, organized by concern:

Directory / File	What it does
main.tsx	Entry point — bootstraps the CLI and renders the terminal UI
cli/	Argument parsing, flag handling, session init
commands/	Slash command implementations (/help, /clear, /compact, etc.)
tools/	Built-in tool implementations (Bash, Read, Write, Edit, Glob, Grep, etc.)
hooks/	Hook event system — PreToolUse, PostToolUse, Stop, Notification, etc.
skills/	Bundled skill loading and discovery
plugins/	Plugin system — loading, scoping, credential management
agents/ / coordinator/	Multi-agent orchestration and task delegation
context/	Context window management, compaction, summarization
state/	Session state, conversation history, memory
screens/	Terminal UI screens (main REPL, dialogs, etc.)
server/	Remote control / API server mode
keybindings/	Keyboard shortcut handling
vim/	Vim mode implementation
voice/	Voice input handling
cost-tracker.ts	Token usage and cost tracking
QueryEngine.ts	Core model query pipeline
schemas/	Zod schemas for tool inputs, hook payloads, config

Notable internals

• Perfetto tracing: Set CLAUDE_CODE_PERFETTO_TRACE=1 to record a Chrome DevTools flame chart of your session. Open in chrome://tracing.
• Hook system: Hooks can do more than allow/deny — PermissionRequest hooks can return updatedPermissions to rewrite the entire permission system. PostToolUse hooks can return updatedMCPToolOutput to intercept MCP responses before the model sees them.
• Sandboxing: Uses Apple seatbelt (macOS) and bubblewrap/seccomp (Linux) with runtime-generated kernel profiles.
• OpenTelemetry: Full OTEL tracing with 16 span types including cost, token, commit, and PR counters.

Source

Extracted from the published npm package: npm install -g @anthropic-ai/claude-code