$ claude --list-feature-flags --include-hidden

Hidden Features

44 feature flags found in the source. Some shipped, some built but flagged off, some internal only. A catalog of what Claude Code can do — and what it's about to do.

Shipped Unshipped Internal Only /buddy Easter Egg

10 shipped

11 unshipped

8 internal

1 easter egg

## Shipped Features

Currently available in production Claude Code

Bash Execution

SHIPPED

BashTool — the "crown jewel." Full shell access with safety layers, destructive command warnings, and preapproved lists.

File Tools

SHIPPED

FileReadTool, FileEditTool, FileWriteTool, GlobTool, GrepTool — the core file manipulation suite.

Agent Tool + Sub-agents

SHIPPED

AgentTool with 5 built-in sub-agents: exploreAgent, planAgent, generalPurposeAgent, verificationAgent, claudeCodeGuideAgent.

Web Search & Fetch

SHIPPED

WebSearchTool and WebFetchTool for live web access during sessions.

MCP Integration

SHIPPED

MCPTool, McpAuthTool, ListMcpResourcesTool, ReadMcpResourceTool — full Model Context Protocol support.

Task Management

SHIPPED

TaskCreateTool, TaskGetTool, TaskListTool, TaskOutputTool, TaskStopTool, TaskUpdateTool for background task orchestration.

Notebook Editing

SHIPPED

NotebookEditTool for Jupyter notebook manipulation — read cells, edit code, view outputs.

Planning Mode

SHIPPED

EnterPlanModeTool / ExitPlanModeV2Tool — structured planning before execution with plan/act separation.

Skills System

SHIPPED

SkillTool — bundled and user-defined skills for repeatable workflows. Slash command integration.

Session Management

SHIPPED

Resume, compact, export, share — full session lifecycle with transcript persistence and replay.

## Unshipped Features

Built but flagged off — 20+ features waiting to go live

Background Agents(Kairos)

UNSHIPPED

24/7 agents with GitHub webhooks, push notifications, and PR monitoring. Agents run continuously on remote servers, watching repos and responding to events.

implications

Turns Claude Code from a session tool into a persistent CI/CD teammate. Could replace simple GitHub Actions workflows.

completeness~85%

Multi-Agent Orchestration(Coordinator Mode)

UNSHIPPED

One Claude spawning worker Claudes with restricted toolsets and scratchpads. Parent agent delegates subtasks and aggregates results.

implications

Parallel work on complex tasks. One agent plans, others execute in isolated contexts. Could dramatically speed up large refactors.

completeness~70%

Browser Control

UNSHIPPED

Full Playwright integration for browser automation — not just web_fetch. Navigate, click, fill forms, take screenshots.

implications

E2E testing, web scraping, visual verification all from within Claude Code sessions. Closes the gap with browser-based AI tools.

completeness~90%

Ultraplan

UNSHIPPED

30-minute Opus session on a remote server for deep task planning. Generates comprehensive implementation plans before execution begins.

implications

Dedicated planning phase with the most capable model. Could produce higher-quality architectural decisions for complex projects.

completeness~60%

Cron Scheduling

UNSHIPPED

CronCreateTool, CronDeleteTool, CronListTool — schedule agents to run on a cron schedule. Basically a CI/CD agent.

implications

Automated code review, dependency updates, test runs on schedule. "Run this agent every morning at 9am" becomes possible.

completeness~80%

Voice Mode(Tengu)

UNSHIPPED

Full push-to-talk voice interface. Uses Deepgram Nova 3 for transcription. Gemstone codenames for feature flags: tengu_cobalt_frost, tengu_amber_quartz.

implications

Hands-free coding, pair programming by talking. Voice commands for navigation and execution.

completeness~75%

Persistent Memory(memdir)

UNSHIPPED

Cross-session memory without external storage. Memory directory system with relevance scoring and age-based decay.

implications

Claude remembers project context, coding preferences, and past decisions across sessions. No more re-explaining your codebase.

completeness~80%

Worktrees

UNSHIPPED

EnterWorktreeTool / ExitWorktreeTool for git worktree isolation. Work on multiple branches simultaneously in separate directories.

implications

Parallel feature development without branch switching. Each agent can work in its own worktree.

completeness~85%

Sleep/Resume Agents

UNSHIPPED

Agents can sleep and self-resume without user prompts. Pause work, wait for external events, then continue autonomously.

implications

Long-running workflows that span hours or days. Wait for CI, wait for review, then resume.

completeness~65%

Plugin System

UNSHIPPED

DiscoverPlugins, ManagePlugins, BrowseMarketplace, AddMarketplace, PluginSettings, ValidatePlugin — a full marketplace ecosystem.

implications

Third-party extensions for Claude Code. Community-built tools, custom workflows, shared configurations.

completeness~70%

Team Tools

UNSHIPPED

TeamCreateTool, TeamDeleteTool for creating and managing multi-agent teams with shared context.

implications

Formalized multi-agent collaboration. Named teams with persistent membership and shared scratchpads.

completeness~50%

## Internal Only

Anthropic employee tools — not accessible in public builds

/ant-trace

INTERNAL

Internal distributed tracing for debugging agent execution paths

/insights

INTERNAL

Usage analytics and session telemetry dashboard for Anthropic engineers

/mock-limits

INTERNAL

Simulate rate limits and quota exhaustion for testing error handling

/reset-limits

INTERNAL

Reset rate limit counters during development and testing

/heapdump

INTERNAL

Capture V8 heap snapshots for memory leak investigation

/perf-issue

INTERNAL

Performance profiling tool for identifying slow tool executions

/debug-tool-call

INTERNAL

Inspect raw tool call payloads, token counts, and execution timing

/stats

INTERNAL

Internal statistics: token usage, tool call frequency, error rates

## /buddy Easter Egg

The ASCII pet system hidden in Claude Code — salt: "friend-2026-401"

An entire gacha pet system built into a coding tool. 18 species, rarity tiers from common to legendary, shiny variants, wearable hats, and stat tracking. Species names are hex-encoded in the source to dodge the internal build scanner. The salt "friend-2026-401" confirms it was an April Fools feature.

species

rarity tiers

legendary drop rate

Species Pool

Names stored as hex in source (e.g., 0x6475636b = "duck") to avoid build scanner detection

duckcapybaradragonghostaxolotlchonkpenguincatdogfrogowlfoxbunnyslimecrabbatbeemushroom

Gacha Rarity Tiers

Common

60%

Uncommon

25%

Rare

10%

Epic

Legendary

Customization

shiny variants

Rare alternate colorways for each species. Visual-only, no stat bonus.

hats

crownwizardpropellertinyduck

Pet Stats

DEBUGGING

How helpful your buddy is at finding bugs

CHAOS

Tendency to suggest unconventional solutions

SNARK

Sass level in buddy comments and reactions

// The buddy subsystem is 6 files, uses hex-encoded species names to avoid the internal build scanner flagging "fun" code, and includes the salt "friend-2026-401" — a reference to April 1st, 2026. Good enough for picking ducks.