Genuinely hidden, gated, or underdocumented capabilities found in the source code — things the public docs don’t cover.

Based on direct source inspection. “Hidden” = hidden from --help, feature-flagged, or dependent on non-public backends. Source presence does not guarantee your build/account has access.

Table of Contents

1. Hidden CLI Flags

These flags are registered with .hideHelp() — they work but won’t appear in claude --help:

Flag What It Does
--teleport Upload local git state to a remote Claude Code session on claude.ai
--remote Create a new remote session (comment in code: “undocumented until GA”)
--remote-control / --rc Enter bridge mode — control from claude.ai web UI (requires BRIDGE_MODE)
--sdk-url Connect to a custom SDK URL for direct-connect sessions
--channels Register for MCP inbound push notifications (KAIROS builds)
--dangerously-load-development-channels Bypass MCP channel allowlist
--enable-auto-mode AI classifier-driven auto-permission (requires TRANSCRIPT_CLASSIFIER)
--advisor <model> Attach a secondary reviewer model
--cowork Switch plugin commands to internal cowork marketplace
--agent-id, --team-name, --teammate-mode, --agent-type Swarm identity flags for multi-agent coordination

Deprecated aliases (still work): --afk and --dangerously-skip-permissions-with-classifiers → map to --enable-auto-mode

Correction: --voice is NOT a CLI flag. Voice is activated via /voice slash command or the voiceEnabled setting. --brief and --proactive are not hidden — they appear in help when their feature flags are on.

2. Feature-Gated Slash Commands

These commands exist but are conditionally registered or hidden:

Command What It Does Gate
/buddy Hatch and interact with a Tamagotchi-style AI pet BUDDY flag + date gate
/voice Toggle hold-to-talk voice dictation VOICE_MODE + Anthropic OAuth
/advisor [model\|off] Attach/detach a secondary reviewer model GrowthBook tengu_sage_compass
/fast [on\|off] Toggle fast inference mode Available when fast mode is supported
/dream Manually trigger memory consolidation Auto-memory must be enabled
/brief Toggle brief/checkpoint mode KAIROS or KAIROS_BRIEF
/ultraplan Launch a remote 30-minute planning session ULTRAPLAN feature flag
/heapdump Dump JavaScript heap to ~/Desktop Always registered, hidden
/thinkback 2025 Claude Code year-in-review stats GrowthBook tengu_thinkback
/remote-control (alias /rc) Enter bridge mode BRIDGE_MODE

3. The Buddy System — A Full Tamagotchi Pet

A fully implemented virtual companion that lives beside your input box.

Activate: /buddy

  • Deterministic generation from your userId via Mulberry32 PRNG — same user = same companion across all devices
  • 18 species: duck, goose, blob, cat, dragon, octopus, owl, penguin, turtle, snail, ghost, axolotl, capybara, cactus, robot, rabbit, mushroom, chonk
  • 5 rarity tiers: common (60%), uncommon (25%), rare (10%), epic (4%), legendary (1%)
  • 6 eye styles, 8 hats (commons get no hat), 1% shiny chance
  • 5 stats: DEBUGGING, PATIENCE, CHAOS, WISDOM, SNARK — one peak stat, one dump stat, floors scale with rarity
  • Soul generation: On first hatch, Claude generates a unique name and personality, stored permanently
  • ASCII sprite animation: 3 frames per species at 500ms tick — idle, fidget, and rare blink frames
  • Speech bubbles (10 seconds, 3-second fade) and /buddy pet hearts animation (2.5 seconds)
  • Anti-cheat: Only the soul persists — bones (species/rarity/stats) are regenerated from userId hash every load

Release window: Teaser April 1-7 2026 (local dates for rolling timezone buzz). Live permanently from April 2026 onward. Always on for internal builds.

4. KAIROS — Persistent Autonomous Assistant Mode

Activate: --assistant flag (feature-gated: KAIROS)

A complete alternate UX where Claude becomes a long-lived autonomous agent persisting across sessions:

  • Append-only daily logs at ~/.claude/projects/<slug>/memory/logs/YYYY/MM/YYYY-MM-DD.md
  • 15-second blocking budget — any command exceeding 15s is auto-backgrounded
  • Proactive <tick> prompts — periodic check-ins where Claude decides what to do next or calls Sleep
  • Brief mode — all output through SendUserMessage tool (structured markdown + attachments + status), not free-form text
  • Exclusive tools: SendUserFile, PushNotification, SubscribePR (GitHub webhook subscriptions), SleepTool
  • Midnight boundary handling — flushes transcript on date change so the dream process can find it
  • Nightly dreaming uses a separate disk-skill variant (distinct from the auto-dream system below)

5. Auto-Dream — Background Memory Consolidation

Runs automatically in the background — no user action needed.

A forked subagent reviews your recent sessions and consolidates learnings into structured memory files.

4-phase process:

  1. Orient — ls memory dir, read index, skim existing topics
  2. Gather — Check daily logs, find drifted memories, grep transcripts narrowly
  3. Consolidate — Write/update memory files, merge duplicates, convert relative dates to absolute
  4. Prune — Update MEMORY.md index (max ~25KB), remove stale pointers

Gates (cheapest checks first):

  • Time: 24+ hours since last consolidation
  • Sessions: 5+ sessions since last consolidation
  • Lock: no other process mid-consolidation
  • Scan throttle: session scanning limited to every 10 minutes

Safety: Bash restricted to read-only commands. Users can kill from the background tasks dialog (Shift+Down).

User control: autoDreamEnabled setting overrides the GrowthBook gate tengu_onyx_plover.

6. Magic Docs — Self-Maintaining Documentation

Files with a # MAGIC DOC: <title> first line are automatically updated by a background agent.

How to use:

  1. Create a markdown file with # MAGIC DOC: My Topic as the first line
  2. Optionally add italic instructions on the next line
  3. Make sure Claude reads that file during a session
  4. A constrained background agent will update it with new learnings

The agent can only edit that specific file — it cannot modify other files. This is triggered by file format, not a command.

7. ULTRAPLAN — Remote 30-Minute Planning Sessions

Activate: Type “ultraplan” in your message (keyword detection) or use /ultraplan

Farms out complex exploration to a remote Claude Code instance (CCR):

  1. Remote session created with plan mode pre-configured
  2. CLI polls every 3 seconds for up to 30 minutes
  3. Remote Claude explores, plans, and calls ExitPlanMode when ready
  4. You approve or reject the plan in the browser (claude.ai)
  5. Rejected plans loop back for iteration
  6. On approval: choose “remote” (execute in cloud) or “teleport to terminal” (execute locally)

Smart keyword detection avoids false positives — skips occurrences inside quotes, paths, identifiers, and questions.

8. Coordinator Mode — Multi-Agent Swarms

Activate: CLAUDE_CODE_COORDINATOR_MODE=1

Transforms Claude into a multi-agent orchestrator:

  • Master coordinator spawns workers via AgentTool in parallel
  • Workers report back as XML <task-notification> messages with status, summary, result, and token usage
  • Coordinator never polls — push-based completion notifications
  • Workers get isolated scratch directories (via tengu_scratch gate) for cross-worker knowledge
  • System prompt enforces: “Never write ‘based on your findings’ — synthesize yourself”
  • 4-phase workflow: Research (parallel) → Synthesis (coordinator) → Implementation → Verification
  • Explicit continue-vs-spawn guidance based on context overlap

9. Speculation — Predictive Response Generation

While you’re still typing, Claude Code speculatively starts generating a response.

  • File writes go to an overlay filesystem (not your real files)
  • If your actual input matches the speculation boundary → overlay committed instantly
  • If it doesn’t match → discarded silently
  • Feature gate: tengu_chomp_inflection (GrowthBook)
  • Result: noticeably lower perceived latency for predictable follow-ups

10. The Advisor Model System

Activate: /advisor <model> or --advisor <model>

Attaches a secondary reviewer/advisor model as a server-side tool:

  • Main model (e.g., Sonnet) can call a stronger model (e.g., Opus) for review
  • Full conversation history is forwarded when the advisor is invoked
  • Beta header: advisor-tool-2026-03-01
  • Does NOT work on Bedrock/Vertex (they don’t support the advisor beta header)
  • Disable with /advisor off or CLAUDE_CODE_DISABLE_ADVISOR_TOOL
  • GrowthBook gate: tengu_sage_compass

11. Voice Mode

Activate: /voice slash command or voiceEnabled setting

  • Hold-to-talk dictation (default keybinding: hold space)
  • Requires Anthropic OAuth — not available with API keys, Bedrock, or Vertex
  • Uses the voice_stream endpoint on claude.ai
  • Multiple audio backends: native audio, SoX, arecord
  • Protected by GrowthBook kill-switch (tengu_amber_quartz_disabled) for emergency off

12. Team Memory Sync

Feature-gated: TEAMMEM

Memories split into private (per-user) and team (shared) directories:

  • Team memory synced to server APIs across authenticated org members
  • Secret scanning prevents leaking sensitive data into shared memory
  • Optimistic locking for conflict resolution
  • Team memory lives at .../memory/team/MEMORY.md
  • Requires first-party OAuth and org-scoped server APIs

13. Remote Triggers — Scheduled Cloud Agents

Tool: RemoteTrigger (deferred, discovered via ToolSearchTool)

Create and manage scheduled remote Claude Code agents via CCR API:

  • create — Schedule a trigger with cron expression
  • list / get — View triggers
  • update — Modify a trigger
  • run — Manually fire a trigger

Requires Claude.ai OAuth. Feature gate: tengu_surreal_dali. Beta: ccr-triggers-2026-01-30.

14. Direct Connect — cc:// Session URLs

Activate: claude server (when DIRECT_CONNECT is enabled)

Creates shareable session URLs that external clients can connect to:

  • WebSocket-based protocol with SDK message format
  • Permission requests forwarded to connecting client
  • Supports interrupt/cancel signals
  • claude open cc://... connects to an existing session (described as internal)
  • Enables custom IDE integrations and frontends

15. Bridge Mode & Remote Control

Activate: --remote-control / --rc (when BRIDGE_MODE compiled in)

Persistent WebSocket connection between local CLI and claude.ai web interface:

  • Use Claude Code from a browser while tools execute locally
  • Exponential backoff reconnection (2s → 2min cap → 10min give-up)
  • JWT token refresh, trusted device enrollment
  • Permission request mediation between web UI and local CLI
  • 31 files in the bridge subsystem — essentially a full RPC framework

Related settings:

  • remoteControlAtStartup — auto-start bridge
  • taskCompleteNotifEnabled, inputNeededNotifEnabled, agentPushNotifEnabled — push notification controls

16. SSH Remote Execution

Activate: claude ssh <host> [dir] (when SSH_REMOTE is enabled)

  • Deploys Claude Code binary to a remote Linux host over SSH
  • API auth tunnels back through the local machine — no separate remote auth setup
  • Build-time feature-gated

17. MCP Channels — Inbound Push Notifications

Activate: --channels plugin:name@marketplace

Register sessions for real-time event delivery from approved MCP servers/plugins:

  • Allowlist controlled by GrowthBook
  • --dangerously-load-development-channels bypasses the allowlist
  • Enables event-driven workflows (e.g., GitHub events, CI notifications)

18. AFK / Auto-Permission Mode

Activate: --enable-auto-mode (hidden flag, requires TRANSCRIPT_CLASSIFIER)

Classifier-assisted automatic permission decisions when the user is away:

  • AI model evaluates each permission request based on conversation context
  • Beta header: afk-mode-2026-01-31
  • Deprecated aliases: --afk, --dangerously-skip-permissions-with-classifiers
  • Designed for unattended, long-running agent workflows

19. Background Sessions (Detached)

Feature-gated: BG_SESSIONS

Run Claude Code sessions in the background with management commands:

  • claude --bg / claude --background — Start a detached session
  • claude ps — List running background sessions
  • claude logs <id> — View session logs
  • claude attach <id> — Reattach to a session
  • claude kill <id> — Stop a session

20. “While You Were Away” Session Recaps

When you return to a session after being away, Claude can generate a short recap card summarizing what happened.

  • Implemented in services/awaySummary.ts
  • Produces short re-entry summaries automatically
  • UI-side feature, not a slash command

21. Tool-Use Summary Generation

For SDK/mobile surfaces, raw tool batches are automatically converted into compact high-level progress summaries:

  • Implemented in services/toolUseSummary/toolUseSummaryGenerator.ts
  • Generates short labels for completed tool batches
  • Used by SDK to provide progress updates to clients and mobile-like single-line rows

22. Auto-Memory Extraction

A background agent automatically extracts memories from your conversations:

  • Runs at end of each complete query loop (when EXTRACT_MEMORIES is enabled)
  • Uses a forked agent that shares the prompt cache
  • Scans existing memory files first to avoid duplicates
  • When the main agent writes memories directly, the background extractor skips that range
  • Writes to ~/.claude/projects/<path>/memory/
  • Auto-memory is enabled by default unless disabled via CLAUDE_CODE_DISABLE_AUTO_MEMORY, --bare, or settings

23. Prompt Suggestions & Follow-Up Generation

Activate: CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION=1

After Claude finishes a response, it can suggest what to ask next:

  • Feature gate: tengu_chomp_inflection (GrowthBook)
  • Env var can force on/off, otherwise GrowthBook + interactive-session checks apply
  • Tied to the speculation system — may pre-generate the speculative response too

24. Deferred Tool Discovery

~18 of Claude Code’s 60+ tools are not sent to the model in every request. They’re discovered on-demand:

  1. Model calls ToolSearchTool with a keyword query
  2. Matching deferred tools’ schemas are returned
  3. Model calls the discovered tool in the same turn

Query syntax:

  • select:TaskCreate,LSP — Direct selection by name
  • task create — Keyword search against names, descriptions, and hints
  • +slack send — Require “slack” in tool name

Why it matters: Keeps the base prompt under 200K tokens. Without this, 60+ tool schemas would consume too much context.

25. Hidden Keybindings

Feature-gated keybindings that only appear when their features are enabled:

Key Combo Action Gate
Space (hold) Push-to-talk voice input VOICE_MODE
Ctrl+Shift+B Toggle Brief mode KAIROS / KAIROS_BRIEF
Ctrl+Shift+F / Cmd+Shift+F Global search QUICK_SEARCH
Ctrl+Shift+P / Cmd+Shift+P Quick open QUICK_SEARCH
Meta+J Toggle terminal panel TERMINAL_PANEL
Shift+Up Message actions menu MESSAGE_ACTIONS
Ctrl+Shift+O Toggle teammate preview Teams enabled

Always-available but often unknown:

Key Combo Action
Ctrl+X Ctrl+K Kill all running agents
Ctrl+_ or Ctrl+Shift+- Undo
Ctrl+X Ctrl+E or Ctrl+G Open external editor
Ctrl+S Stash current input
Meta+P Model picker
Meta+O Fast mode toggle
Meta+T Thinking mode toggle
Ctrl+E Toggle permission explanation (in confirmation dialog)
Ctrl+D Toggle permission debug info (in confirmation dialog)

All keybindings are overridable via ~/.claude/keybindings.json.

26. Lesser-Known Environment Variables

Debug & profiling:

Variable Purpose
CLAUDE_CODE_PROFILE_STARTUP=1 Full startup profiling with memory snapshots
CLAUDE_CODE_PROFILE_QUERY=1 Profile query pipeline timing
CLAUDE_CODE_DEBUG_REPAINTS=1 Show component owner chain for every terminal repaint
CLAUDE_CODE_PERFETTO_TRACE Enable Perfetto tracing format
CLAUDE_CODE_TERMINAL_RECORDING Record terminal in asciinema format
CLAUDE_CODE_COMMIT_LOG=/path Log slow renders for profiling

Behavioral overrides:

Variable Purpose
CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC Disable ALL non-essential network traffic (most restrictive privacy)
CLAUDE_CODE_DISABLE_AUTO_MEMORY Disable automatic memory management
CLAUDE_CODE_DISABLE_CRON Disable cron job scheduler
CLAUDE_CODE_DISABLE_FILE_CHECKPOINTING Disable file snapshot backups
CLAUDE_CODE_DISABLE_ADVISOR_TOOL Disable advisor model
CLAUDE_CODE_ENABLE_PROMPT_SUGGESTION Enable speculative next-prompt suggestions
CLAUDE_CODE_MAX_OUTPUT_TOKENS Override max output tokens per response
CLAUDE_CODE_MAX_CONTEXT_TOKENS Override max context window
CLAUDE_CODE_EFFORT_LEVEL Set effort: low\|medium\|high\|max\|auto
CLAUDE_CODE_COORDINATOR_MODE=1 Enable coordinator mode
CLAUDE_CODE_SIMPLE Same as --bare — skip hooks, LSP, plugins, background tasks
CLAUDE_CONFIG_DIR Override ~/.claude config directory
CLAUDE_ENV_FILE Path to env file to source on startup

Provider switching:

Variable Purpose
CLAUDE_CODE_USE_BEDROCK=1 Use AWS Bedrock
CLAUDE_CODE_USE_VERTEX=1 Use Google Vertex AI
CLAUDE_CODE_USE_FOUNDRY=1 Use Azure Foundry
CLAUDE_CODE_CLIENT_CERT / CLAUDE_CODE_CLIENT_KEY mTLS client certificates

Prompt caching control:

Variable Purpose
DISABLE_PROMPT_CACHING Disable globally
DISABLE_PROMPT_CACHING_HAIKU / _SONNET / _OPUS Disable per model
DISABLE_INTERLEAVED_THINKING Disable thinking blocks

27. Lesser-Known Settings Keys

In ~/.claude/settings.json — things most people don’t know you can set:

Key What It Does
apiKeyHelper External command to fetch API key (e.g., 1password read op://vault/key)
awsCredentialExport Command to export AWS credentials for Bedrock
env Arbitrary environment variables injected into every session
effortLevel Default effort level: low\|medium\|high\|max\|auto
alwaysThinkingEnabled Force extended thinking on every request
spinnerVerbs Custom spinner verb list
spinnerTipsOverride Custom tip messages during spinner
worktree.symlinkDirectories Directories to symlink in worktrees (saves disk)
worktree.sparsePaths Git sparse-checkout paths for monorepo worktrees
autoMemoryDirectory Custom path for memory storage
autoDreamEnabled Enable/disable auto-dream consolidation
minSleepDurationMs Minimum SleepTool duration
skipWebFetchPreflight Skip WebFetch URL validation
disableBypassPermissionsMode Prevent entering bypass mode
allowManagedPermissionRulesOnly Enterprise: only admin-defined permission rules
allowManagedMcpServersOnly Enterprise: only admin-defined MCP servers
allowManagedHooksOnly Enterprise: only admin-defined hooks
allowedHttpHookUrls URL allowlist for HTTP hooks
httpHookAllowedEnvVars Env vars HTTP hooks can interpolate
remote.defaultEnvironmentId Default remote environment for CCR
minimumVersion Enforce minimum Claude Code version

28. CLAUDE.md Loading — Hidden Discovery Paths

Beyond the standard project CLAUDE.md, Claude Code loads instructions from:

  • ~/.claude/CLAUDE.md — User-level global instructions
  • Parent directories up to git root — all CLAUDE.md files in parent dirs are included
  • .claude/CLAUDE.md — Inside the .claude directory
  • .claude/rules/*.md — Per-project rule files (all .md files in this directory)
  • @include-style references inside memory files

29. Internal-Only Commands

Registered only when USER_TYPE === 'ant' — not in public builds:

backfill-sessions, break-cache, bughunter, ctx_viz, good-claude, init-verifiers, force-snip, mock-limits, bridge-kick, subscribe-pr, reset-limits, share, ant-trace, perf-issue, env, oauth-refresh, debug-tool-call, agents-platform, autofix-pr, onboarding

30. Build-Time Feature Flags

Compile-time flags via feature() from bun:bundle. When off, code is eliminated entirely from the binary:

Flag Feature
COORDINATOR_MODE Multi-agent coordinator
VOICE_MODE Voice input
KAIROS / KAIROS_BRIEF Persistent assistant mode
PROACTIVE Autonomous mode
BRIDGE_MODE Remote control bridge
SSH_REMOTE SSH remote execution
DIRECT_CONNECT cc:// URL handling
BG_SESSIONS Background sessions
TEMPLATES Template/new/reply flows
TEAMMEM Team memory sync
TRANSCRIPT_CLASSIFIER AI permission classification
BUDDY Tamagotchi pet
ULTRAPLAN Remote planning
EXTRACT_MEMORIES Auto-memory extraction
WORKFLOW_SCRIPTS Workflow automation
QUICK_SEARCH Quick search (Ctrl+Shift+F)
TERMINAL_PANEL Terminal panel (Meta+J)
MESSAGE_ACTIONS Message action menu (Shift+Up)
CONTEXT_COLLAPSE Context collapse optimization
HISTORY_SNIP History snipping
MCP_SKILLS MCP skill discovery
DAEMON Long-running daemon

Compiled from static analysis of the Claude Code source. Corrections applied from the audited version. Features behind flags may not be in your build.