Undocumented Features in Claude Code

TL;DR

What this document covers: A comprehensive inventory of 88+ feature flags, hidden CLI commands, undocumented environment variables, and obscure internal systems found in the Claude Code source code. These are not covered in the official documentation and many are gated behind Anthropic's internal build system.

Key undocumented patterns:

88 compile-time feature flags via feature() from bun:bundle - most are dead-code eliminated in public builds
Fast-path subcommands handled before Commander parsing (remote-control, daemon, ps, attach, kill, ssh)
Hidden CLI flags (--bare, --prefill, --agents, --fork-session, --effort)
60+ environment variables for feature enable/disable, operational config, and internal Anthropic use
24 ant-only commands gated behind USER_TYPE=ant
Anti-debug protection - process exits if debugger detected (non-ant builds)
Secret scanner - 23+ detection rules for client-side secret scanning

Why this matters: Understanding these internals helps you:

Discover hidden capabilities
Debug feature availability issues
Configure advanced options
Understand the full scope of Claude Code's architecture

mindmap
  root((Undocumented<br/>Features))
    Feature Flags
      88 compile-time flags
      KAIROS multi-agent
      PROACTIVE agent
      BRIDGE_MODE remote
      CACHED_MICROCOMPACT
    Hidden Commands
      Fast-path subcommands
      Ant-only commands
      Feature-gated commands
    Environment Variables
      Feature enable/disable
      Operational config
      Internal/Anthropic
    Secret Systems
      Anti-debug protection
      Secret scanner
      Deep link handler

Build-Time Feature Flags
Hidden CLI Flags
Fast-Path Subcommands
Anthropic-Only Commands
Feature-Gated Commands
Undocumented Environment Variables
Notable Hidden Systems

Build-Time Feature Flags

Claude Code uses compile-time feature flags via feature() (imported from bun:bundle) for dead code elimination. There are 88 unique flags. Most are undocumented and gated behind Anthropic's internal build configuration.

How Feature Flags Work

// From source: src/services/kairos/index.ts
if (feature('KAIROS')) {
  // This code is tree-shaken if KAIROS flag is false
  const kairosModule = require('./kairosModule')
  kairosModule.init()
}

Key insight: These are build-time, not runtime flags. Once the binary is built, features are either included or dead-code eliminated. You cannot enable them via environment variables in public builds.

flowchart LR
    A[Source Code] --> B{Feature Flag}
    B -->|Enabled| C[Included in Build]
    B -->|Disabled| D[Dead Code Elimination]
    C --> E[Binary with Feature]
    D --> F[Binary without Feature]

Agent Orchestration

Flag	File References	Description	Status
`KAIROS`	Multiple	Full "assistant" agent mode with multi-agent capabilities	ant-only
`KAIROS_BRIEF`	commands/brief.tsx	KAIROS brief sub-feature	ant-only
`KAIROS_CHANNELS`	services/	Channel-based notifications for KAIROS	ant-only
`KAIROS_DREAM`	services/	KAIROS dream feature (purpose unclear)	ant-only
`KAIROS_GITHUB_WEBHOOKS`	commands/subscribe-pr.tsx	GitHub webhook subscriptions	ant-only
`KAIROS_PUSH_NOTIFICATION`	services/	Push notification support	ant-only
`PROACTIVE`	components/, hooks/	Agent acts without user prompts	ant-only
`COORDINATOR_MODE`	coordinator/	Multi-agent orchestrator	ant-only
`AGENT_TRIGGERS`	services/	Cron-based agent scheduling	ant-only
`AGENT_TRIGGERS_REMOTE`	tools/	Remote trigger tool	ant-only
`AGENT_MEMORY_SNAPSHOT`	memdir/	Persistent memory snapshots	ant-only
`VERIFICATION_AGENT`	tools/	Separate verification agent	ant-only
`BUILTIN_EXPLORE_PLAN_AGENTS`	tools/	Built-in explore/plan agent pair	ant-only
`FORK_SUBAGENT`	commands/fork.ts	Fork a sub-agent from current session	ant-only
`TRANSCRIPT_CLASSIFIER`	hooks/	Auto permission mode from conversation analysis	ant-only

KAIROS Architecture:

flowchart TB
    subgraph KAIROS["KAIROS Multi-Agent System"]
        direction TB
        K1[Coordinator] --> K2[Explore Agent]
        K1 --> K3[Plan Agent]
        K1 --> K4[Code Agent]
        K1 --> K5[Verify Agent]
        
        K2 --> K6[Channel Notifications]
        K3 --> K6
        K4 --> K6
        K5 --> K6
        
        K6 --> K7[GitHub Webhooks]
        K6 --> K8[Push Notifications]
    end
    
    subgraph PROACTIVE["PROACTIVE Mode"]
        P1[Monitors Context] --> P2[Acts Without Prompts]
    end

Remote & Daemon

Flag	Description	Status
`BRIDGE_MODE`	Remote control over WebSocket	ant-only
`DAEMON`	Long-running daemon supervisor	ant-only
`BG_SESSIONS`	Background session management (`ps`, `logs`, `attach`, `kill`)	ant-only
`DIRECT_CONNECT`	`cc://` and `cc+unix://` URL scheme support	ant-only
`LODESTONE`	Deep link URI handling	ant-only
`SSH_REMOTE`	`claude ssh <host>` sessions	ant-only
`CCR_REMOTE_SETUP`	CCR remote setup web command	ant-only
`CCR_AUTO_CONNECT`	CCR auto-connect	ant-only
`CCR_MIRROR`	CCR session mirroring	ant-only
`SELF_HOSTED_RUNNER`	CI self-hosted runner	ant-only
`BYOC_ENVIRONMENT_RUNNER`	Bring-your-own-cloud runner	ant-only
`UDS_INBOX`	Unix domain socket inbox/peers	ant-only

Remote Control Architecture:

flowchart LR
    subgraph Local["Local Machine"]
        L1[Claude Code CLI]
        L2[WebSocket Client]
    end
    
    subgraph Remote["Remote/Cloud"]
        R1[CCR Server]
        R2[WebSocket Server]
        R3[Session Mirror]
    end
    
    subgraph Browser["Browser"]
        B1[claude.ai/code]
    end
    
    L1 <-->|BRIDGE_MODE| L2
    L2 <-->|WebSocket| R2
    R2 --> R1
    R2 --> R3
    B1 <-->|HTTPS| R1

Model & API

Flag	Description	Status
`ANTI_DISTILLATION_CC`	Content filtering against distillation	ant-only
`CACHED_MICROCOMPACT`	Cached micro-compaction (cache-preserving)	ant-only
`PROMPT_CACHE_BREAK_DETECTION`	Detect prompt cache breaks	ant-only
`CONTEXT_COLLAPSE`	Context compaction	ant-only
`REACTIVE_COMPACT`	Reactive compaction	ant-only
`TOKEN_BUDGET`	API-side token budgets	ant-only
`UNATTENDED_RETRY`	Unattended retry logic	ant-only
`BASH_CLASSIFIER`	Bash command classification	ant-only
`CONNECTOR_TEXT`	Connector text summarization	ant-only
`EFFORT_BETA_HEADER`	Effort level beta	ant-only

Tools & Integrations

Flag	Description	Status
`CHICAGO_MCP`	Computer-use MCP (screen/keyboard control)	ant-only
`MCP_SKILLS`	MCP server skill fetching	ant-only
`MCP_RICH_OUTPUT`	Rich MCP output rendering	ant-only
`EXPERIMENTAL_SKILL_SEARCH`	Skill search/discovery	ant-only
`SKILL_IMPROVEMENT`	Skill improvement surveys	ant-only
`RUN_SKILL_GENERATOR`	Skill generator	ant-only
`WORKFLOW_SCRIPTS`	Workflow tool/scripts	ant-only
`WEB_BROWSER_TOOL`	Web browser tool	ant-only
`TOOL_SEARCH`	Tool search capability	ant-only
`TREE_SITTER_BASH`	Tree-sitter bash parsing	ant-only
`POWERSHELL_AUTO_MODE`	PowerShell auto-mode detection	ant-only

UI & UX

Flag	Description	Status
`BUDDY`	Buddy companion sprite with notifications	ant-only
`AUTO_THEME`	Auto theme detection	ant-only
`HISTORY_PICKER`	History picker dialog	ant-only
`HISTORY_SNIP`	Force-snip history command	ant-only
`MESSAGE_ACTIONS`	Message action menus	ant-only
`QUICK_SEARCH`	Quick search feature	ant-only
`TERMINAL_PANEL`	Terminal panel	ant-only
`STREAMLINED_OUTPUT`	Streamlined output mode	ant-only
`VOICE_MODE`	Voice interaction	ant-only

Telemetry & Internal

Flag	Description	Status
`PERFETTO_TRACING`	Perfetto-compatible trace export	ant-only
`ENHANCED_TELEMETRY_BETA`	Enhanced telemetry	ant-only
`SHOT_STATS`	Shot statistics (ant-only)	ant-only
`SLOW_OPERATION_LOGGING`	Slow operation logging	ant-only
`ABLATION_BASELINE`	Harness-science experiments (disables thinking, compaction, memory, tasks)	ant-only
`ULTRAPLAN`	Extended planning mode	ant-only
`ULTRATHINK`	Extended thinking mode	ant-only
`TORCH`	Torch feature	ant-only

Hidden CLI Flags

Undocumented top-level flags from src/main.tsx:

Core Flags

Flag	Description	Example
`--bare`	Stripped-down hermetic auth mode	`claude --bare`
`--json-schema <schema>`	Enforce JSON output schema	`claude --json-schema '{...}'`
`--include-hook-events`	Emit hook events in output	`claude --include-hook-events`
`--input-format <format>`	Structured input parsing	`claude --input-format json`
`--replay-user-messages`	Replay mode	`claude --replay-user-messages`
`--prefill <text>`	Prefill the assistant response	`claude --prefill "I'll help..."`
`--agents <json>`	Pass agent definitions as JSON	`claude --agents '[...]'`
`--setting-sources <sources>`	Custom setting sources	`claude --setting-sources ...`
`--disable-slash-commands`	Disable all slash commands	`claude --disable-slash-commands`

Session Management

Flag	Description	Example
`--fork-session`	Fork current session	`claude --continue --fork-session`
`--from-pr <pr>`	Start from a PR	`claude --from-pr 123`
`--agent <agent>`	Launch a specific agent	`claude --agent explore`
`--session-id <uuid>`	Force a specific session UUID	`claude --session-id ...`
`--name <name>`	Name the session	`claude --name "Bug Fix"`

Model & API

Flag	Description	Example
`--betas <betas>`	Pass API beta flags	`claude --betas prompt-caching-2024-07-31`
`--effort <level>`	Set reasoning effort level	`claude --effort high`
`--fallback-model <model>`	Fallback model on primary failure	`claude --fallback-model sonnet`
`--file <specs>`	Pre-load file specs	`claude --file "src/*.ts"`

Integration

Flag	Description	Example
`--strict-mcp-config`	Strict MCP config mode	`claude --strict-mcp-config`
`--plugin-dir <path>`	Plugin directory path	`claude --plugin-dir ./plugins`
`--chrome/--no-chrome`	Chrome integration toggle	`claude --chrome`
`--debug-file <path>`	Debug log to file	`claude --debug-file debug.log`

Fast-Path Subcommands

Handled before Commander argument parsing in src/entrypoints/cli.tsx:

flowchart LR
    A[CLI Input] --> B{Fast-Path?}
    B -->|Yes| C[Handle Immediately]
    B -->|No| D[Commander Parsing]
    C --> E[Execute]
    D --> E

Remote & Session Management

Subcommand	Aliases	Description	Example
`remote-control`	`rc`, `remote`, `sync`	Remote session control	`claude rc`
`daemon`	—	Long-running daemon mode	`claude daemon`
`ps`	—	List background sessions	`claude ps`
`logs`	—	View session logs	`claude logs <id>`
`attach`	—	Attach to running session	`claude attach <id>`
`kill`	—	Kill a session	`claude kill <id>`
`new`	—	New session	`claude new`
`list`	—	List sessions	`claude list`
`reply`	—	Reply to a session	`claude reply <id>`

CI/CD Runners

Subcommand	Description	Example
`environment-runner`	CI environment runner	`claude environment-runner`
`self-hosted-runner`	CI self-hosted runner	`claude self-hosted-runner`

Connectivity

Subcommand	Description	Example
`ssh`	SSH remote sessions	`claude ssh production`
`open`	Internal `cc://` deep link handler	`claude open cc://...`

Anthropic-Only Commands

Commands gated behind USER_TYPE === 'ant' (Anthropic employees):

Functional Commands (7 work)

Command	Status	Description
`commit`	✅ Works	Creates git commit with generated message
`commit-push-pr`	✅ Works	Commit, push, and create PR
`init-verifiers`	✅ Works	Generate verification skills
`version`	✅ Works	Print build version
`tag`	✅ Works	Toggle searchable session tags
`files`	✅ Works	List files in context
`bridge-kick`	⚠️ Partial	Requires active bridge connection

Stubbed Commands (17 don't work)

Even with USER_TYPE=ant, these export { isEnabled: () => false }:

// All 17 stubs look like this:
export default { isEnabled: () => false, isHidden: true, name: 'stub' };

Command	Inferred Purpose
`backfill-sessions`	Backfill session data for analytics
`break-cache`	Break/toggle prompt cache behavior
`bughunter`	Code review via bughunter system
`ctx_viz`	Context visualization
`good-claude`	Feedback/rating mechanism
`issue`	Create/interact with GitHub issues
`mock-limits`	Mock rate limit scenarios
`reset-limits`	Reset rate limits
`onboarding`	User onboarding flow
`share`	Share session data
`summary`	Summarize conversation
`teleport`	Remote session teleportation
`ant-trace`	Internal tracing/debugging
`perf-issue`	Report performance issues
`env`	Set session-scoped environment variables
`oauth-refresh`	Refresh OAuth tokens
`debug-tool-call`	Debug tool call execution
`agents-platform`	⚠️ Missing module (would crash)
`autofix-pr`	Auto-fix PR review comments

Feature-Gated Commands

Commands available only when their corresponding feature flag is enabled:

flowchart TB
    subgraph Commands["Feature-Gated Commands"]
        C1[remote-setup] --> F1[CCR_REMOTE_SETUP]
        C2[fork] --> F2[FORK_SUBAGENT]
        C3[buddy] --> F3[BUDDY]
        C4[proactive] --> F4[PROACTIVE]
        C5[brief] --> F5[KAIROS]
        C6[assistant] --> F6[KAIROS]
        C7[bridge] --> F7[BRIDGE_MODE]
        C8[voice] --> F8[VOICE_MODE]
        C9[workflows] --> F9[WORKFLOW_SCRIPTS]
        C10[torch] --> F10[TORCH]
    end

Command	Feature Flag	Description
`remote-setup`	`CCR_REMOTE_SETUP`	Remote setup wizard
`fork`	`FORK_SUBAGENT`	Fork sub-agent
`buddy`	`BUDDY`	Buddy companion
`proactive`	`PROACTIVE` or `KAIROS`	Proactive agent
`brief`	`KAIROS` or `KAIROS_BRIEF`	KAIROS brief
`assistant`	`KAIROS`	KAIROS assistant
`bridge`	`BRIDGE_MODE`	Bridge control
`remote-control-server`	`DAEMON` + `BRIDGE_MODE`	Remote control server
`voice`	`VOICE_MODE`	Voice interaction
`peers`	`UDS_INBOX`	Unix domain socket peers
`workflows`	`WORKFLOW_SCRIPTS`	Workflow management
`torch`	`TORCH`	Torch feature
`subscribe-pr`	`KAIROS_GITHUB_WEBHOOKS`	PR webhook subscriptions
`ultraplan`	`ULTRAPLAN`	Extended planning
`force-snip`	`HISTORY_SNIP`	Force history snip

Undocumented Environment Variables

Feature Enable (22 variables)

Variable	Description	Default
`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS`	Multi-agent swarm spawning	`false`
`CLAUDE_CODE_ENABLE_CFC`	Claude-in-Chrome integration	`false`
`CLAUDE_CODE_ENABLE_XAA`	Extended Authorization Architecture (MCP OAuth)	`false`
`ENABLE_AGENT_SWARMS`	Agent swarm support	`false`
`ENABLE_TOOL_SEARCH`	Tool search/discovery	`auto`
`ENABLE_SESSION_PERSISTENCE`	Session persistence	`false`
`ENABLE_BETA_TRACING_DETAILED`	Detailed beta tracing	`false`
`ENABLE_PROMPT_CACHING_1H_BEDROCK`	1-hour prompt caching (Bedrock)	`false`
`ENABLE_ENHANCED_TELEMETRY_BETA`	Enhanced telemetry beta	`false`
`ENABLE_PID_BASED_VERSION_LOCKING`	PID-based version locking	`false`
`ENABLE_LOCKLESS_UPDATES`	Lockless update mechanism	`false`
`ENABLE_LSP_TOOL`	LSP tool	`false`
`ENABLE_CLAUDEAI_MCP_SERVERS`	claude.ai MCP servers	`false`
`ENABLE_MCP_LARGE_OUTPUT_FILES`	Large MCP output files	`false`
`ENABLE_CLAUDE_CODE_SM_COMPACT`	Session-memory compaction	`false`

Feature Disable (37 variables)

Variable	Description	Default
`CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS`	Kill switch for all experimental beta headers	`false`
`CLAUDE_CODE_DISABLE_FILE_CHECKPOINTING`	File checkpoint/rollback	`false`
`CLAUDE_CODE_DISABLE_AUTO_MEMORY`	Automatic memory extraction	`false`
`CLAUDE_CODE_DISABLE_MESSAGE_ACTIONS`	Message action menus	`false`
`CLAUDE_CODE_DISABLE_BACKGROUND_TASKS`	Background task execution	`false`
`CLAUDE_CODE_DISABLE_VIRTUAL_SCROLL`	Virtual scrolling	`false`
`CLAUDE_CODE_DISABLE_ADVISOR_TOOL`	Advisor tool	`false`
`CLAUDE_CODE_DISABLE_POLICY_SKILLS`	Policy-based skills	`false`
`CLAUDE_CODE_DISABLE_SM_COMPACT`	Session-memory compaction	`false`
`CLAUDE_CODE_DISABLE_TERMINAL_TITLE`	Terminal title setting	`false`
`CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY`	Feedback surveys	`false`
`CLAUDE_CODE_DISABLE_LEGACY_MODEL_REMAP`	Legacy model remapping	`false`
`CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC`	Non-essential network traffic	`false`
`CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK`	Non-streaming fallback	`false`
`CLAUDE_CODE_DISABLE_GIT_INSTRUCTIONS`	Git instructions	`false`
`CLAUDE_CODE_DISABLE_OFFICIAL_MARKETPLACE_AUTOINSTALL`	Marketplace auto-install	`false`
`CLAUDE_CODE_DISABLE_CLAUDE_MDS`	Claude MDS	`false`
`CLAUDE_CODE_DISABLE_PRECOMPACT_SKIP`	Pre-compact skip	`false`
`DISABLE_PROMPT_CACHING`	All prompt caching	`false`
`DISABLE_PROMPT_CACHING_HAIKU`	Prompt caching for Haiku	`false`
`DISABLE_PROMPT_CACHING_SONNET`	Prompt caching for Sonnet	`false`
`DISABLE_PROMPT_CACHING_OPUS`	Prompt caching for Opus	`false`
`DISABLE_COMPACT`	Compaction entirely	`false`
`DISABLE_AUTO_COMPACT`	Auto-compaction	`false`
`DISABLE_INTERLEAVED_THINKING`	Interleaved thinking	`false`
`DISABLE_AUTOUPDATER`	Auto-updater	`false`
`DISABLE_LOGIN_COMMAND`	`/login` command	`false`
`DISABLE_LOGOUT_COMMAND`	`/logout` command	`false`
`DISABLE_FEEDBACK_COMMAND`	Feedback command	`false`
`DISABLE_DOCTOR_COMMAND`	`/doctor` command	`false`
`DISABLE_INSTALL_GITHUB_APP_COMMAND`	GitHub app install	`false`
`DISABLE_UPGRADE_COMMAND`	`/upgrade` command	`false`
`DISABLE_EXTRA_USAGE_COMMAND`	Extra-usage command	`false`
`DISABLE_COST_WARNINGS`	Cost warnings	`false`
`DISABLE_ERROR_REPORTING`	Error reporting	`false`
`DISABLE_TELEMETRY`	Telemetry	`false`
`DISABLE_INSTALLATION_CHECKS`	Installation checks	`false`

Internal / Anthropic (16 variables)

Variable	Description	Access
`USER_TYPE`	Set to `'ant'` for internal Anthropic employees	ant-only
`CLAUDE_INTERNAL_FC_OVERRIDES`	GrowthBook feature flag overrides	ant-only
`CLAUDE_CODE_GB_BASE_URL`	GrowthBook base URL	ant-only
`USE_LOCAL_OAUTH`	Local OAuth config	ant-only
`USE_STAGING_OAUTH`	Staging OAuth config	ant-only
`CLAUDE_CODE_CUSTOM_OAUTH_URL`	Custom OAuth URL	ant-only
`CLAUDE_CODE_ABLATION_BASELINE`	Ablation baseline config	ant-only
`CLAUDE_ENABLE_STREAM_WATCHDOG`	Stream health watchdog	ant-only
`IS_DEMO`	Demo environment	ant-only
`CLAUDE_CODE_COWORKER_TYPE`	Coworker type for telemetry	ant-only
`CLAUDE_CODE_TAGS`	Session tags	ant-only
`MCP_CLIENT_SECRET`	MCP OAuth client secret	ant-only
`MCP_XAA_IDP_CLIENT_SECRET`	MCP XAA IdP client secret	ant-only
`CLAUDE_CODE_SIMPLE`	Bare/simple mode flag	ant-only

Operational (13 variables)

Variable	Description	Example
`CLAUDE_CODE_REMOTE`	Running in CCR environment	`1`
`CLAUDE_CODE_ENTRYPOINT`	Entrypoint identifier	`cli`, `sdk-cli`, `mcp`, `local-agent`
`CLAUDE_CODE_ACTION`	Running as GitHub Action	`1`
`CLAUDE_CODE_SESSION_ACCESS_TOKEN`	Remote session access token	`token`
`CLAUDE_CODE_HOST_PLATFORM`	Host platform override	`linux`, `darwin`, `win32`
`CLAUDECODE`	Set to `'1'` for child processes	`1`
`ANTHROPIC_CUSTOM_HEADERS`	Custom HTTP headers	`Name: Value\nName2: Value2`
`ANTHROPIC_CUSTOM_MODEL_OPTION`	Pre-validated model name	`sonnet`
`ANTHROPIC_BETAS`	Comma-separated extra beta headers	`prompt-caching-2024-07-31,extended-output`
`CLAUDE_CODE_EXTRA_BODY`	Extra JSON body params	`'{"key": "value"}'`
`CLAUDE_CODE_EXTRA_METADATA`	Extra metadata	`'{"source": "ci"}'`
`API_TIMEOUT_MS`	API timeout in milliseconds	`600000` (10 min)
`CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`	Override auto-compact threshold percentage	`80` (compact at 80% of context window)

Notable Hidden Systems

Anti-Debug Protection

File: src/main.tsx:266-271

In non-ant builds, the process exits if a debugger is detected:

if (process.env.USER_TYPE !== 'ant' && isDebuggerDetected()) {
  console.error('Debugger detected. Exiting.')
  process.exit(1)
}

Purpose: Prevent reverse engineering of production builds.

Secret Scanner

File: src/services/teamMemorySync/secretScanner.ts

Implements client-side secret scanning before team memory uploads, with 23+ detection rules:

const SECRET_PATTERNS = [
  { name: 'AWS Access Key', pattern: /AKIA[0-9A-Z]{16}/ },
  { name: 'GitHub Token', pattern: /gh[pousr]_[A-Za-z0-9_]{36}/ },
  { name: 'Slack Token', pattern: /xox[baprs]-[0-9a-zA-Z-]+/ },
  { name: 'Private Key', pattern: /-----BEGIN (RSA |DSA |EC |OPENSSH )?PRIVATE KEY-----/ },
  // ... 20+ more patterns
]

Why client-side? Prevents secrets from ever reaching the server.

Deep Link Handler

Files: src/services/lodestone/, src/services/directConnect/

cc:// and cc+unix:// URL scheme support via the LODESTONE and DIRECT_CONNECT feature flags:

// Example deep link
cc://open?path=/home/user/project&session=abc-123

// Unix domain socket
cc+unix:///var/run/claude.sock

Plugin System

Directory: src/plugins/

Full plugin infrastructure with:

/plugin command - Manage plugins
/reload-plugins command - Reload plugin code
--plugin-dir flag - Custom plugin directory
Plugin manifest schema (plugin.json)
Hooks, MCP servers, LSP servers

Vim Mode

Directory: src/vim/

Complete vim keybinding support:

Text objects (words, paragraphs, quotes)
Operators (delete, yank, change)
Motions (h/j/k/l, w/b/e, 0/^/$, gg/G)
Visual mode (character, line, block)
Command mode (:w, :q, :s/foo/bar/)

Memdir System

Directory: src/memdir/

Persistent memory directory for cross-session memory:

user/ - Private user preferences
feedback/ - Approach corrections
project/ - Team-shared project state
reference/ - External system pointers

Buddy Companion

Directory: src/buddy/

A UI sprite/companion system with notifications:

Animated sprite in terminal
Notification badges
Status indicators
Easter egg interactions

experimentalSystemReminder

File: src/Tool.ts:275

An experimental field on agent definitions that re-injects a system reminder every user turn:

experimentalSystemReminder?: string  // Re-injected every turn

Purpose: Keep agent focused on specific behavior throughout long conversations.

Documentation: docs/experimental-system-reminder.md

Summary

Claude Code contains 88+ feature flags, 60+ environment variables, and numerous hidden commands and systems. Most are:

Build-time gated - Dead-code eliminated if flag is false
ant-only - Only available to Anthropic employees
Undocumented - Not mentioned in official docs

The sheer scope of these features indicates Claude Code is a much larger platform than the public CLI suggests, with infrastructure for:

Multi-agent orchestration (KAIROS)
Remote/distributed execution (CCR, Bridge, SSH)
CI/CD integration (Runners, GitHub Actions)
Voice interaction (Voice Mode)
Browser integration (Chrome, CFC)
Advanced security (Secret Scanner, Anti-Debug)

based on alleged Claude Code source analysis (src/main.tsx, src/entrypoints/cli.tsx, src/commands.ts, src/services/, src/utils/permissions/)

FilesExpand file tree

undocumented-features.md

Latest commit

History

undocumented-features.md

File metadata and controls

Undocumented Features in Claude Code

TL;DR

Table of Contents

Build-Time Feature Flags

How Feature Flags Work

Agent Orchestration

Remote & Daemon

Model & API

Tools & Integrations

UI & UX

Telemetry & Internal

Hidden CLI Flags

Core Flags

Session Management

Model & API

Integration

Fast-Path Subcommands

Remote & Session Management

CI/CD Runners

Connectivity

Anthropic-Only Commands

Functional Commands (7 work)

Stubbed Commands (17 don't work)

Feature-Gated Commands

Undocumented Environment Variables

Feature Enable (22 variables)

Feature Disable (37 variables)

Internal / Anthropic (16 variables)

Operational (13 variables)

Notable Hidden Systems

Anti-Debug Protection

Secret Scanner

Deep Link Handler

Plugin System

Vim Mode

Memdir System

Buddy Companion

experimentalSystemReminder

Summary