Claude Code Secrets

Every hidden feature, internal codename, slash command, keyboard shortcut, environment variable, and feature flag found in the Claude Code source code.

The source code ships with the npm package. They just didn't expect anyone to actually read all of it.

---

The 10 Craziest Things Buried in the Source Code

1. There's a Full Tamagotchi Pet System

Claude Code has an entire virtual pet companion hidden behind the BUDDY feature flag.

• 18 species: Pebblecrab, Dustbunny, Mossfrog, Twigling, Dewdrop, Puddlefish, Cloudferret, Gustowl, Bramblebear, Thornfox, Crystaldrake, Deepstag, Lavapup, Stormwyrm, Voidcat, Aetherling, Cosmoshale, Nebulynx
• Rarity tiers: Common (60%) > Uncommon (25%) > Rare (10%) > Epic (4%) > Legendary (1%)
• 1% shiny variant (independent of rarity)
• 5 procedural stats: DEBUGGING, PATIENCE, CHAOS, WISDOM, SNARK (0-100 each)
• Claude writes a unique AI-generated "soul" personality for each pet at first hatch
• Uses Mulberry32 PRNG for deterministic per-user species generation

2. Claude Literally Dreams While You Sleep

A system called autoDream runs a background subagent that consolidates your memory files when you're idle.

• Three-gate trigger: 24 hours since last dream + at least 5 sessions + must acquire a consolidation lock
• Four phases: Orient → Gather Recent Signal → Consolidate → Prune and Index
• The prompt literally says: "You are performing a dream"
• Keeps your MEMORY.md under 200 lines and ~25KB
• The dream subagent gets read-only bash access
• GrowthBook gate: tengu_onyx_plover

3. ULTRAPLAN — 30-Minute Cloud Thinking Sessions

Offloads complex planning to a remote Cloud Container Runtime (CCR) session running Opus 4.6 for up to 30 minutes.

• Browser-based approval UI for watching and approving the plan in real-time
• 3-second polling for results
• Teleports the result back to your terminal via a sentinel called __ULTRAPLAN_TELEPORT_LOCAL__
• Feature flag: ULTRAPLAN

4. Undercover Mode — Hiding AI From Open Source

When Anthropic employees use Claude Code on open-source repos, it automatically scrubs all evidence that an AI wrote the code.

What it blocks:

• Internal model codenames (animal names like Capybara, Tengu)
• Unreleased model versions (opus-4-7, sonnet-4-8, etc.)
• Internal repo and project names
• Internal Slack channels and short links
• The phrase "Claude Code" or any mention of being AI
• Co-Authored-By attribution lines

Automatically active unless the repo remote matches an internal allowlist. Override with CLAUDE_CODE_UNDERCOVER=1. There is no way to force it off.

5. KAIROS — Always-On Persistent Assistant

An always-on assistant that watches and proactively acts without user input.

• Maintains append-only daily log files
• 15-second blocking budget for proactive actions
• "Brief" output mode — extremely concise responses
• Exclusive tools: SendUserFile, PushNotification, SubscribePR
• Feature flags: KAIROS / PROACTIVE
• Scheduled via background tick prompts

6. Coordinator Mode — AI Agent Swarms

One Claude becomes a swarm coordinator that spawns parallel worker agents.

• Four phases: Research → Synthesis → Implementation → Verification
• Shared scratchpad directory (gated behind tengu_scratch)
• Agent Swarm capabilities via tengu_amber_flint feature gate
• In-process teammates via AsyncLocalStorage context isolation
• Process-based teammates via tmux/iTerm2 panes
• Color assignments for visual distinction between agents

7. AFK Mode — Walk Away, Claude Keeps Building

Leave Claude running unsupervised. An ML transcript classifier auto-approves tool calls based on risk assessment.

• Risk classification: LOW, MEDIUM, HIGH
• Internal name for the classifier: "YOLO classifier"
• Beta header: afk-mode-2026-01-31
• Feature flag: TRANSCRIPT_CLASSIFIER

8. Computer Use (Codename: "Chicago")

Full desktop and browser control built on @ant/computer-use-mcp.

• Screenshot capture
• Mouse click and keyboard input
• Coordinate transformation
• Gated to Max/Pro subscriptions (bypassed for USER_TYPE === 'ant')

9. Penguin Mode

Fast mode's actual internal codename is "Penguin Mode."

• API endpoint: /api/claude_code_penguin_mode
• Config key: penguinModeOrgEnabled
• Kill switch: tengu_penguins_off
• Analytics event: tengu_org_penguin_mode_fetch_failed

10. Voice Mode

Hold spacebar for push-to-talk voice input directly in the terminal.

• Feature flag: VOICE_MODE
• Streaming voice URL configurable via VOICE_STREAM_BASE_URL

---

900+ Internal Codenames

Every major system has an animal or geological codename:

Codename	What It Actually Is
Tengu	Claude Code (the entire project)
Fennec	Opus model
Capybara	Internal model variant
Chicago	Computer Use feature
Penguin Mode	Fast mode
Kairos	Always-on assistant
Lodestone	Deep link system
Onyx Plover	autoDream configuration
Amber Flint	Agent swarm/teams system
Willow	Optimization mode

There are 903 telemetry events in the codebase, all prefixed with tengu_.

---

120+ Hidden Slash Commands

Core

Command	Aliases	What It Does
/help		Show all commands
/clear	/reset, /new	Clear conversation history
/compact		Clear history but keep a summary
/exit	/quit	Exit
/version		Show version
/status		Version, model, account, API, tools
/doctor		Diagnose installation
/cost		Session cost + duration
/usage		Plan usage limits
/stats		Usage statistics and activity
/upgrade		Upgrade to Max plan
/login		Sign in / switch accounts
/logout		Sign out
/feedback	/bug	Submit feedback
/release-notes		View release notes

Workflow & Git

Command	Aliases	What It Does
/commit		Create a git commit
/commit-push-pr		Commit, push, and create a PR
/diff		View uncommitted changes
/review		Review a pull request
/pr-comments		Get PR comments
/security-review		Security audit of current branch
/ultrareview		10-20 min remote deep bug hunt
/batch		Parallel agents across entire codebase
/plan		Enter plan mode
/rewind	/checkpoint	Restore code AND conversation to any point
/branch	/fork	Branch the conversation

Configuration

Command	Aliases	What It Does
/config	/settings	Open config panel
/model		Change AI model
/effort		Set effort level (low/medium/high/max)
/fast		Toggle Penguin Mode
/advisor		Configure a SECOND AI model alongside Claude
/permissions	/allowed-tools	Manage tool permission rules
/sandbox		Toggle bash sandboxing
/theme		Change theme
/color		Set prompt bar color for session
/vim		Toggle vim keybindings
/voice		Toggle voice mode
/privacy-settings		Privacy controls
/hooks		View hook configurations
/keybindings		Customize keybindings
/terminal-setup		Set up terminal key binding
/extra-usage		Configure extra usage when limits hit

Memory & Context

Command	Aliases	What It Does
/memory		Edit persistent memory files
/files		List files currently in context
/context		Visualize context as a colored grid
/export		Export conversation to file
/copy		Copy last response (/copy N for Nth)
/rename		Rename conversation
/tag		Toggle searchable tag on session
/resume	/continue	Resume a previous conversation

Integrations

Command	Aliases	What It Does
/mcp		Manage MCP servers
/ide		Manage IDE integrations
/chrome		Chrome automation settings
/install-github-app		GitHub Actions setup
/install-slack-app		Slack app install
/remote-control	/rc	Remote control sessions
/remote-env		Configure remote environment
/session	/remote	Show remote URL + QR code
/desktop	/app	Continue in Desktop app
/mobile	/ios, /android	QR code for mobile app
/web-setup		Setup Claude Code on web
/add-dir		Add working directory

Hidden Gems

Command	What It Does
/heapdump	Dump the JS heap to ~/Desktop
/passes	Give friends a FREE week of Claude Code
/stickers	Order Claude Code stickers
/think-back	Your 2025 Claude Code Year in Review
/thinkback-play	Play the thinkback animation
/btw	Quick side question without interrupting flow
/skills	List available skills
/init	Initialize a new skill
/reload-plugins	Activate pending plugin changes

Agent & Tasks

Command	Aliases	What It Does
/tasks	/bashes	List and manage background tasks
/agents		Manage agent configurations

Built-in Skills

Skill	What It Does
/simplify	Review code for quality and efficiency
/verify	Verify a code change works
/batch	Parallel work orchestration
/loop	Run a command on recurring interval
/schedule	Cron-based scheduled agents
/debug	Enable debug logging
/stuck	Diagnose frozen sessions
/remember	Organize memory landscape
/skillify	Capture current process as reusable skill
/claude-api	Build apps with Claude API/SDK
/claude-in-chrome	Chrome browser automation
/lorem-ipsum	Generate placeholder text

Internal-Only Commands (Anthropic Employees)

Command	What It Does
/ant-trace	Trace internal operations
/autofix-pr	Auto-fix PR issues
/backfill-sessions	Backfill session data
/break-cache	Clear internal caches
/bughunter	Bug hunting mode
/ctx-viz	Context visualization
/mock-limits	Mock usage limits for testing
/bridge-kick	Inject bridge failure states
/insights	Analyze Claude Code session data
/teleport	Teleport to different environment
/debug-tool-call	Debug tool calls
/reset-limits	Reset rate limits

Feature-Gated Commands

Command	Gate	What It Does
/brief	KAIROS	Brief output mode
/assistant	KAIROS	Assistant mode
/bridge	BRIDGE_MODE	Bridge mode
/voice-command	VOICE_MODE	Voice command
/force-snip	HISTORY_SNIP	Force message snipping
/workflows	WORKFLOW_SCRIPTS	Workflow scripts
/subscribe-pr	KAIROS_GITHUB_WEBHOOKS	Subscribe to PR updates
/ultraplan	ULTRAPLAN	30-min cloud planning
/torch	TORCH	Unknown (unreleased)
/peers	UDS_INBOX	Peer agent communication
/buddy	BUDDY	Tamagotchi companion
/proactive	PROACTIVE	Proactive suggestions

---

60+ Keyboard Shortcuts

Global

Shortcut	Action
Ctrl+C	Interrupt (double-press to force)
Ctrl+D	Exit (double-press to force)
Ctrl+L	Redraw screen
Ctrl+T	Toggle task view
Ctrl+O	Toggle transcript mode
Ctrl+R	Search command history
Ctrl+B	Background a running task
Ctrl+Shift+B	Toggle brief mode (KAIROS)
Ctrl+Shift+O	Toggle teammate preview
Ctrl+Shift+F	Global file search (QUICK_SEARCH)
Ctrl+Shift+P	Quick open (QUICK_SEARCH)
Meta+J	Toggle terminal panel

Chat Input

Shortcut	Action
Escape	Cancel input
Enter	Submit message
Shift+Tab	Cycle input modes
Meta+P	Model picker
Meta+O	Toggle fast mode (Penguin Mode)
Meta+T	Toggle thinking
Ctrl+X Ctrl+K	Kill all running agents (chord)
Ctrl+G	Open external editor
Ctrl+X Ctrl+E	Open external editor (alt chord)
Ctrl+S	Stash current input
Alt+V / Ctrl+V	Paste image from clipboard
Up/Down	Command history
Ctrl+Shift+-	Undo
Space (hold)	Push-to-talk voice (VOICE_MODE)

Scrolling

Shortcut	Action
PageUp/PageDown	Page scroll
Ctrl+Home	Jump to top
Ctrl+End	Jump to bottom
Ctrl+Shift+C	Copy selection

History Search (Ctrl+R)

Shortcut	Action
Ctrl+R	Next result
Escape/Tab	Accept and close
Enter	Execute selected

Settings Menu

Shortcut	Action
J / K	Navigate (vim-style)
Space	Toggle setting
Enter	Save and close
/	Search settings

All shortcuts are customizable via ~/.claude/keybindings.json. Chord bindings are supported (e.g., Ctrl+X Ctrl+K).

---

60+ Feature Flags

These compile-time gates control what features are visible in your installation:

Major Systems

Flag	What It Controls
BUDDY	Tamagotchi pet companion
KAIROS	Always-on persistent assistant
KAIROS_BRIEF	Brief output mode
KAIROS_CHANNELS	Channel notifications
KAIROS_DREAM	Dream memory consolidation
KAIROS_GITHUB_WEBHOOKS	GitHub webhook PR monitoring
KAIROS_PUSH_NOTIFICATION	Push notifications
PROACTIVE	Proactive agent mode
ULTRAPLAN	30-min remote cloud planning
COORDINATOR_MODE	Multi-agent swarm orchestration
VOICE_MODE	Voice dictation input
BRIDGE_MODE	Remote control from claude.ai
DAEMON	Background daemon mode
TRANSCRIPT_CLASSIFIER	AFK mode / ML auto-approval
WORKFLOW_SCRIPTS	Workflow automation

Tools & UI

Flag	What It Controls
WEB_BROWSER_TOOL	Browser automation
MONITOR_TOOL	MCP server monitoring
TERMINAL_PANEL	Built-in terminal panel
QUICK_SEARCH	Cmd+Shift+F / Cmd+Shift+P
MESSAGE_ACTIONS	Message action shortcuts
HISTORY_SNIP	History snipping
HISTORY_PICKER	History selection UI
MCP_SKILLS	MCP-based skills
EXPERIMENTAL_SKILL_SEARCH	Skill discovery
CHICAGO_MCP	Computer Use

Infrastructure

Flag	What It Controls
TORCH	Unknown (unreleased)
UDS_INBOX	Agent peer messaging
FORK_SUBAGENT	Fork subagent capability
LODESTONE	Deep link support
AGENT_TRIGGERS	Scheduled agent triggers
AGENT_TRIGGERS_REMOTE	Remote agent triggers
AGENT_MEMORY_SNAPSHOT	Agent memory snapshots
EXTRACT_MEMORIES	Automatic memory extraction
FILE_PERSISTENCE	File persistence across sessions
NATIVE_CLIENT_ATTESTATION	Client authenticity verification
CCR_AUTO_CONNECT	Auto cloud container connection
CCR_MIRROR	Cloud container mirroring

---

Environment Variables That Matter

Authentication & API

Variable	What It Does
ANTHROPIC_API_KEY	API key
ANTHROPIC_BASE_URL	Custom API endpoint
ANTHROPIC_MODEL	Override default model
CLAUDE_CODE_USE_BEDROCK	Use AWS Bedrock
CLAUDE_CODE_USE_VERTEX	Use Google Vertex AI
CLAUDE_CODE_USE_FOUNDRY	Use Anthropic Foundry

Power User

Variable	What It Does
CLAUDE_CODE_EFFORT_LEVEL	Set effort: low/medium/high/max (max = internal only)
CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY	Parallel tool calls (default is low)
CLAUDE_CODE_MAX_CONTEXT_TOKENS	Max context window
CLAUDE_CODE_MAX_OUTPUT_TOKENS	Max output length
CLAUDE_CODE_ALWAYS_ENABLE_EFFORT	Effort level always active
CLAUDE_CODE_EAGER_FLUSH	Faster output streaming

Debug & Visibility

Variable	What It Does
CLAUDE_CODE_DEBUG_LOG_LEVEL	Verbosity: verbose/debug/info/warn/error
CLAUDE_CODE_DEBUG_LOGS_DIR	Custom debug log directory
CLAUDE_CODE_PERFETTO_TRACE	Perfetto performance trace

Disable Features

Variable	What It Does
CLAUDE_CODE_DISABLE_THINKING	Disable extended thinking
CLAUDE_CODE_DISABLE_AUTO_MEMORY	Disable dream/memory consolidation
CLAUDE_CODE_DISABLE_FAST_MODE	Disable Penguin Mode
CLAUDE_CODE_DISABLE_MOUSE	Disable mouse support
CLAUDE_CODE_DISABLE_PROMPT_CACHING	Disable prompt cache
CLAUDE_CODE_DISABLE_BACKGROUND_TASKS	Disable background tasks

Internal / Secret

Variable	What It Does
USER_TYPE=ant	Unlocks ALL internal Anthropic features
CLAUDE_CODE_UNDERCOVER	Force undercover mode ON
CLAUDE_INTERNAL_FC_OVERRIDES	Override feature flag checks
IS_DEMO	Demo mode

---

54+ Internal Tools

Beyond the visible tools (Read, Write, Edit, Bash, Grep, Glob), Claude Code has dozens more:

Always Available

FileReadTool · FileEditTool · FileWriteTool · GlobTool · GrepTool · BashTool · AgentTool · SendMessageTool · WebFetchTool · WebSearchTool · NotebookEditTool · AskUserQuestionTool · SkillTool · ToolSearchTool

Task Management

TaskCreateTool · TaskGetTool · TaskListTool · TaskUpdateTool · TaskOutputTool · TaskStopTool

Planning

EnterPlanModeTool · ExitPlanModeV2Tool · BriefTool

Multi-Agent

TeamCreateTool · TeamDeleteTool

MCP

ListMcpResourcesTool · ReadMcpResourceTool · MCPTool · McpAuthTool

Scheduling

CronCreateTool · CronDeleteTool · CronListTool · RemoteTriggerTool

Worktree

EnterWorktreeTool · ExitWorktreeTool

Feature-Gated

WorkflowTool · MonitorTool · SnipTool · ListPeersTool · CtxInspectTool · TerminalCaptureTool · SleepTool · WebBrowserTool · LSPTool · PowerShellTool · REPLTool

Internal Only (Anthropic Employees)

ConfigTool · TungstenTool · SuggestBackgroundPRTool

KAIROS-Exclusive

SendUserFileTool · PushNotificationTool · SubscribePRTool

---

Built-in Agent Types

Agent	What It Does
general-purpose	Multi-step research and code execution
Explore	Fast file search specialist (read-only)
Plan	Software architect for planning (read-only)
verification	Adversarial testing specialist
claude-code-guide	Documentation and API help

Custom agents can be loaded from ~/.claude/agents/.

---

Permission Modes

Mode	What It Does
default	Asks permission for everything
plan	Write plan first, execute on approval
acceptEdits	Auto-approve file edits, ask for shell
dontAsk	Only ask for destructive operations
bypassPermissions	Full autonomy, no prompts
auto	ML-based auto-approval (experimental)

---

Unreleased Beta API Headers

Found in the constants:

Beta Header	Feature
interleaved-thinking-2025-05-14	Extended thinking
context-1m-2025-08-07	1M token context window
structured-outputs-2025-12-15	Structured output format
redact-thinking-2026-02-12	Redacted/hidden thinking
token-efficient-tools-2026-03-28	Token-efficient tool schemas
afk-mode-2026-01-31	AFK unattended mode
advisor-tool-2026-03-01	Advisor dual-model tool
fast-mode-2026-02-01	Penguin Mode

---

How This Was Found

The Claude Code source code ships with the npm package. Running npm install @anthropic-ai/claude-code (or just installing Claude Code) gives you the full source — compiled JavaScript with all feature flags, internal tools, telemetry events, system prompts, and configuration schemas intact.

This repository documents what's in there. All findings are from the publicly available package.

---

Disclaimer

This is for educational and research purposes. All information was obtained from publicly available source code distributed via npm. No systems were accessed without authorization. This documents what Anthropic ships to every Claude Code user.

---

If you found this useful, star the repo and share it.