Claude Code's Secret Features

We analyzed a leaked version of the Claude Code source and read every line. What we found is a codebase with an elaborate multi-layer feature gating system that hides 24+ fully-implemented features from public builds.

This is not speculation. Every claim on this site is backed by direct source code references.

The Scale of What’s Hidden

Metric	Count
`process.env.USER_TYPE === 'ant'` checks	291+ across 143+ files
Build-time `feature()` flags via `bun:bundle`	89 compile-time gates
Runtime `tengu_*` GrowthBook flags	500+ server-side feature flags
Major hidden feature systems	24+

When Anthropic builds Claude Code for public release, the Bun bundler evaluates every feature() call and every USER_TYPE check. If the flag is false, dead-code elimination strips the entire code path. The features literally do not exist in the binary you download.

How It Works in Three Sentences

Build time: process.env.USER_TYPE is set to 'ant' only for internal builds. The bundler inlines the value and eliminates unreachable branches.
Compile time: feature('SOME_FLAG') from bun:bundle returns true or false at bundle time. False branches are removed entirely.
Runtime: Even if code survives bundling, GrowthBook feature flags (tengu_*) gate functionality server-side, returning false for non-Anthropic accounts.

The Hidden Features

Here is a summary of the major hidden features we have identified so far:

Feature	Gate	Description
Coordinator Mode	`feature('COORDINATOR_MODE')`	Multi-agent orchestration where a coordinator delegates to parallel workers
Fork Subagent	`feature('FORK_SUBAGENT')`	Cache-optimized child agents that inherit parent conversation history
Agent Swarms	`tengu_amber_flint`	Leader-teammate model with tmux/iTerm pane spawning and shared memory
Kairos	`feature('KAIROS')`	Scheduled background agent execution with cron-like triggers
Auto Dream	`tengu_onyx_plover`	Automated memory consolidation and context summarization
Scratchpad	`tengu_scratch`	Cross-worker durable storage directory for multi-agent workflows
Hearth	`feature('HEARTH')`	Persistent background services and daemon management
MCP Auth	`feature('MCP_AUTH')`	OAuth-based authentication for MCP server connections
Notebook Support	`feature('NOTEBOOK')`	Native Jupyter notebook editing and execution
GitHub PR Integration	`feature('GITHUB_PR')`	Deep pull request review and management workflows
Interop Mode	`feature('INTEROP')`	Cross-tool integration protocol for editor extensions
Prompt Caching	`ant`-gated	Advanced prompt cache management and optimization
Custom Models	`ant`-gated	Internal model selection beyond public offerings
MDM Policies	`ant`-gated	Mobile Device Management policy enforcement
Telemetry Dashboard	`ant`-gated	Internal usage analytics and performance monitoring
Subscription Tiers	`ant`-gated	Feature differentiation across max/pro/plus/basic tiers
Permission Bridge	`ant`-gated	Cross-agent permission synchronization
Team Memory Sync	`ant`-gated	Shared context propagation across agent teams
Monorepo Detection	`MONOREPO_ROOT_DIR`	Special features when running inside Anthropic’s monorepo
Plugin System	`ant`-gated	Extended plugin architecture beyond public MCP
Background Tasks	`ant`-gated	Long-running task management with progress tracking
Session Handoff	`ant`-gated	Transfer active sessions between devices/instances
Smart Retry	`ant`-gated	Intelligent request retry with backoff strategies
Config Sync	`ant`-gated	Cross-device configuration synchronization

Deep Dives

How Anthropic Hides Features in Plain Sight The three-layer gating system: build-time elimination, runtime flags, and auth gates.

Coordinator Mode: The Hidden Multi-Agent Orchestra Multi-worker orchestration, fork subagents, and agent swarms with shared memory.