v2.1.84Reverse Engineered

Claude Code Unleashed

|

A systematic reverse engineering analysis of both the native macOS binary and the npm package, uncovering undocumented systems, feature flags, environment variables, and hidden capabilities.

Explore findingsMethodology
0API Beta Flags
0Feature Flags
0Slash Commands
0Settings
0Env Vars
0Hidden Systems

01

Hidden Systems

Undocumented subsystems discovered inside the Claude Code binary. These are fully-implemented features gated behind server-side flags.

Iron Gate

Enabled

Fail-closed safety for auto-mode. Blocks actions when classifier unavailable.

flag: tengu_iron_gate_closeddefault: true

Willow Mode

Disabled

Idle detection. After 75 min + 100K tokens, shows dialog/hint.

flag: tengu_willow_modedefault: "off"

Kairos

Mixed

Auto-brief + cron scheduling subsystem.

flag: tengu_kairos_brief, tengu_kairos_crondefault: false / true

Auto-Mode Classifier

Active

2-stage LLM (fast + thinking) decides permissions. Temperature=0, 4096 max tokens.

flag: N/Adefault: N/A

Auto-Dream

Server-gated

Background memory consolidation on idle. PID-locked.

flag: tengu_onyx_ploverdefault: null

Fast Mode Cooldown

Active

"Penguin mode" state machine. Tracks cooldown with resetAt/reason.

flag: tengu_penguins_offdefault: null

Agent Teams

Hidden Flag

Multi-agent collaboration via AsyncLocalStorage.

flag: --agent-teams flagdefault: N/A

Teleport

Hidden

Session bundling/transfer between local CLI and cloud.

flag: --teleport flagdefault: N/A

Ultraplan

Hardcoded Disabled

Remote multi-agent planning. 10-30 min Opus planning on web.

flag: isEnabled:()=>!1default: N/A

Session Memory

Disabled

Cross-session persistent memory (separate from CLAUDE.md).

flag: tengu_session_memorydefault: false

Harbor

Disabled

MCP channel push notifications with allowlists.

flag: tengu_harbordefault: false

02

API Beta Flags

30 beta header strings sent with API requests. These unlock experimental API capabilities.

Active (16)

claude-code-20250219Active
interleaved-thinking-2025-05-14Active
context-1m-2025-08-07Active
context-management-2025-06-27Active
oauth-2025-04-20Active
redact-thinking-2026-02-12Active
structured-outputs-2025-12-15Active
web-search-2025-03-05Active
prompt-caching-scope-2026-01-05Active
effort-2025-11-24Active
task-budgets-2026-03-13Active
fast-mode-2026-02-01Active
afk-mode-2026-01-31Active
advisor-tool-2026-03-01Active
advanced-tool-use-2025-11-20Active
tool-search-tool-2025-10-19Active

SDK / Infrastructure (10)

bedrock-2023-05-31SDK
vertex-2023-10-16SDK
ccr-byoc-2025-07-29SDK
ccr-triggers-2026-01-30SDK
environments-2025-11-01SDK
mcp-servers-2025-12-04SDK
files-api-2025-04-14SDK
message-batches-2024-09-24SDK
token-counting-2024-11-01SDK
skills-2025-10-02SDK

Docs-only (4)

experiment-v2Docs
compact-2026-01-12Docs
structured-outputs-2025-11-13Docs
mcp-client-2025-11-20Docs

03

Disabled & Secret Commands

Slash commands that exist in the codebase but are hidden, flag-gated, or hardcoded disabled.

Hardcoded Disabled

/ultraplan

Remote multi-agent planning

/files

Context file listing

/bridge-kick

Debug bridge restart

/version

Show version info

/tag

Session tagging

Flag-gated

/think-back

Year in Review

/thinkback-play

Animated review playback

/web-setup

GitHub connection setup

/schedule

Cron-based agent scheduling

Easter Eggs

/stickers

Order physical stickers!

/dream

Manual memory consolidation

/add-dir/agent-teams/allowed-tools/api-docs/api-key/bridge-kick/bug/clear/compact/config/configure-mcp/cost/debug/diff/doctor/dream/dump/edit/explain/files/fix/forget/help/hooks/ide/ignore/init/install/kick/listen/login/logout/logs/ls/mcp/memory/model/new/notebook/open/outline/patch/permissions/plan/pr/profile/project/quit/read/refresh/release/review/run/schedule/search/send/settings/setup/share/shortcut/skill/spawn/status/stickers/summary/tag/task/terminal/think-back/thinkback-play/think/ultraplan/undo/version/vim/web-setup/worktree

04

Feature Flags

All 79 tengu_* flags extracted from cli.js. Prefixed with "tengu_" -- Anthropic's internal Statsig flag namespace.

tengu_iron_gate_closedHOT

Fail-closed safety when auto-mode classifier is unavailable

default: true
tengu_penguins_offHOT

Server-side kill switch for fast mode with custom reason message

default: null
tengu_willow_modeHOT

Idle session return behavior (dialog/hint after inactivity)

default: "off"
tengu_auto_background_agentsHOT

Enables automatic background task spawning with 2min timeout

default: false
tengu_session_memoryHOT

Persistent memory file across a session, survives compacts

default: false
tengu_kairos_briefHOT

Gates Brief tool and brief-only output mode

default: false
tengu_kairos_cronHOT

Gates cron/scheduled tasks and /loop command

default: true
tengu_onyx_ploverHOT

Auto-dream background memory consolidation config

default: null
tengu_ultraplan_modelHOT

Which model to use for ultraplan generation

default: opus46
tengu_auto_mode_configHOT

Master config for Auto Mode: enabled state, model, classifier, allowModels

default: {}
tengu_sage_compassHOT

Configures Advisor tool: enabled, model, canUserConfigure

default: {}
tengu_amber_flint

Gates Agent Teams (multi-agent swarm) feature

default: true
tengu_amber_prism

Appends correction preference instructions to system prompt

default: false
tengu_amber_quartz_disabled

Kill switch for Remote Control (CCR bridge) feature

default: false
tengu_amber_wren

Configures file read size limits and token caps for tool output

default: {}
tengu_attribution_header

Controls whether attribution header is sent with API requests

default: true
tengu_bramble_lintel

Min conversation turns before memory extraction triggers

default: null
tengu_bridge_system_init

Sends system/init message to Remote Control bridge on connect

default: false
tengu_ccr_bridge

Master gate for Remote Control (CCR) bridge feature

default: false
tengu_chomp_inflection

Enables prompt suggestions feature (inline suggestions after input)

default: false
tengu_chrome_auto_enable

Auto-enables Claude-in-Chrome MCP integration

default: false
tengu_cicada_nap_ms

Cooldown period (ms) for startup prefetch operations

default: 0
tengu_cobalt_frost

Enables Nova 3 speech-to-text engine for voice streaming

default: false
tengu_cobalt_lantern

Enables GitHub token sync and /web-setup for Remote Control

default: false
tengu_collage_kaleidoscope

Uses native C module for clipboard image detection on macOS

default: true
tengu_compact_cache_prefix

Cache-sharing optimization for compact/summarization

default: false
tengu_compact_streaming_retry

Retry logic with exponential backoff for streaming compact failures

default: false
tengu_copper_bridge

Controls WebSocket bridge URL for Chrome MCP communication

default: false
tengu_coral_fern

Adds 'Searching past context' instructions to system prompt

default: false
tengu_destructive_command_warning

Inline warnings for destructive shell commands (git reset, rm -rf)

default: false
tengu_disable_streaming_to_non_streaming_fallback

Disables fallback from streaming to non-streaming on error

default: false
tengu_fgts

Fine-grained tool streaming (eager input streaming)

default: false
tengu_grey_wool

Legacy model name remapping to current equivalents

default: true
tengu_harbor

Master gate for Channels (plugins/extensions marketplace)

default: false
tengu_harbor_permissions

Enables permissions system for harbor/channels plugins

default: false
tengu_hawthorn_steeple

Tool output deduplication/replacement in messages

default: false
tengu_hawthorn_window

Token window size for tool output dedup (default 200K)

default: null
tengu_herring_clock

Team memory directory functionality (shared org memory)

default: false
tengu_immediate_model_command

Makes /model command execute immediately without picker UI

default: false
tengu_jade_anvil_4

Reorders rate-limit menu to nudge toward paid options

default: false
tengu_kairos_cron_config

Configuration for cron/durable scheduled task subsystem

default: {}
tengu_keybinding_customization_release

Enables custom keybinding support in configuration

default: false
tengu_lapis_finch

Auto plugin hint recommendations based on tool usage

default: false
tengu_lodestone_enabled

Auto-registers claude-cli:// deep link protocol handler

default: false
tengu_marble_sandcastle

Restricts fast mode to native binary (Bun) build only

default: false
tengu_miraculo_the_bard

Controls fast mode prefetch strategy (sync vs async)

default: false
tengu_moth_copse

Filters memory messages from extraction prompts (anti-recursion)

default: false
tengu_paper_halyard

Filters Project/Local CLAUDE.md from system prompt

default: false
tengu_passport_quail

Enables memory extraction in non-interactive/SDK sessions

default: false
tengu_permission_explainer

AI-generated risk explanations for tool permission prompts

default: false
tengu_pewter_ledger

Plan mode output style: trim/cut/cap verbosity levels

default: null
tengu_pid_based_version_locking

PID-based version file locking for concurrent instances

default: false
tengu_plan_mode_interview_phase

Interview/clarification phase before plan generation

default: false
tengu_plugin_official_mkt_git_fallback

Git clone fallback when GCS marketplace fetch fails

default: true
tengu_plum_vx3

Lightweight model with disabled thinking for web search

default: false
tengu_quiet_fern

Gate value forwarded to VSCode extension experiments

default: false
tengu_quiet_hollow

Redact-thinking API beta to hide thinking from responses

default: false
tengu_remote_backend

Remote backend session mode with interactive REPL

default: false
tengu_satin_quoll

Per-tool output token limits and MCP tool output caps

default: {}
tengu_slate_heron

Auto tool output trimming based on session time gaps

default: {}
tengu_slate_prism

Agent progress summaries in SDK sessions

default: true
tengu_slate_thimble

Allows memory extraction in fully headless/SDK mode

default: false
tengu_slim_subagent_claudemd

Strips CLAUDE.md from subagent prompts to reduce size

default: true
tengu_sm_compact

Session-memory-aware compaction strategy

default: false
tengu_surreal_dali

Remote Trigger (scheduled remote agent) and /schedule command

default: false
tengu_system_prompt_global_cache

Global cache control on system prompt for API cache hits

default: false
tengu_tern_alloy

Subagent fanout nudge prompt suggestion copy variant

default: "off"
tengu_thinkback

Enables /think-back and /thinkback-play Year in Review

default: false
tengu_tide_elm

Effort-high nudge prompt suggestion copy variant

default: "off"
tengu_timber_lark

/loop command nudge prompt suggestion copy variant

default: "off"
tengu_tool_pear

Deferred tool loading and tool input schema stripping

default: false
tengu_tool_search_unsupported_models

Model patterns that don't support tool search/deferred loading

default: null
tengu_trace_lantern

Detailed beta tracing for diagnostics/debugging

default: false
tengu_turtle_carbon

Master gate for ultrathink/ultraplan extended thinking

default: true
tengu_vscode_cc_auth

Gate for VSCode Claude Code authentication features

default: false
tengu_vscode_onboarding

VSCode extension onboarding configuration

default: {}
tengu_vscode_review_upsell

VSCode extension review/upsell configuration

default: {}
tengu_cobalt_frost_2

Extended Nova 3 STT configuration variant

default: false
tengu_pebble_leaf_prune

Smarter pruning of orphaned transcript message branches

default: false

05

Configuration Settings

All 118 settings across 12 categories. Scoped to user (~/.claude/settings.json), project (.claude/settings.json), managed (enterprise), or local (.claude/settings.local.json).

model

Default model for conversations

user/project
availableModels

List of models available in model picker

managed
modelOverrides

Override model for specific contexts (subagent, compact, etc.)

user/project
advisorModel

Stronger model used by the Advisor tool for review

user/project
effortLevel

Thinking effort level: low, medium, high, or max

user/project
alwaysThinkingEnabled

Force extended thinking on every request

user
fastMode

Enable fast mode for quicker, lighter responses

user
fastModePerSessionOptIn

Require per-session opt-in for fast mode

user
agent

Agent behavior configuration (max turns, model, etc.)

user/project
permissions.allow

Allowlist of tools/commands that run without asking

user/project/managed
permissions.deny

Blocklist of tools/commands that are always denied

user/project/managed
permissions.ask

Tools/commands that always require confirmation

user/project/managed
permissions.defaultMode

Default permission mode: ask, auto, or deny

user/project/managed
permissions.disableBypassPermissionsMode

Prevent users from using --dangerously-skip-permissions

managed
permissions.disableAutoMode

Prevent auto-mode from being enabled

managed
permissions.additionalDirectories

Extra directories the agent can access beyond CWD

user/project
allowManagedPermissionRulesOnly

Only allow permission rules from managed config

managed
sandbox.enabled

Enable sandboxed command execution

user/project/managed
sandbox.failIfUnavailable

Fail if sandbox cannot be created (vs falling back)

managed
sandbox.autoAllowBashIfSandboxed

Auto-allow bash commands when sandboxed

user/project
sandbox.network.allowedDomains

Domains the sandbox can reach

user/project/managed
sandbox.network.allowManagedDomainsOnly

Only allow domains from managed config

managed
sandbox.network.allowUnixSockets

Specific Unix socket paths to allow

user/project
sandbox.network.allowAllUnixSockets

Allow all Unix socket connections

user/project
sandbox.network.allowLocalBinding

Allow binding to localhost ports

user/project
sandbox.filesystem.allowWrite

Paths the sandbox can write to

user/project
sandbox.filesystem.denyWrite

Paths the sandbox cannot write to

user/project/managed
outputStyle

Output formatting style: streaming, markdown, or plain

user
language

UI language / locale override

user
syntaxHighlightingDisabled

Disable syntax highlighting in code blocks

user
spinnerTipsEnabled

Show tips in the loading spinner

user
spinnerVerbs

Custom verbs for the loading spinner animation

user
statusLine

Configure the status line display format

user
defaultView

Default view mode on startup

user
voiceEnabled

Enable voice input/output

user
prefersReducedMotion

Reduce animations and transitions

user
showThinkingSummaries

Show AI thinking block summaries in output

user
terminalTitleFromRename

Update terminal title from /rename command

user
promptSuggestionEnabled

Enable inline prompt suggestions after input

user
autoMemoryEnabled

Automatically extract and store memories from conversations

user
autoMemoryDirectory

Directory for auto-extracted memory files

user
autoDreamEnabled

Background memory consolidation when idle

user
claudeMdExcludes

Glob patterns to exclude from CLAUDE.md discovery

user/project
cleanupPeriodDays

Days before old sessions/logs are cleaned up

user
plansDirectory

Directory for storing generated plans

user/project
enableAllProjectMcpServers

Auto-enable all MCP servers defined in project config

user
enabledPlugins

List of enabled marketplace plugins

user
pluginConfigs

Per-plugin configuration objects

user/project
extraKnownMarketplaces

Additional marketplace URLs beyond the default

user/project/managed
strictKnownMarketplaces

Only allow plugins from known marketplaces

managed
channelsEnabled

Enable the channels/extensions system

user
enabledMcpjsonServers

MCP servers enabled from .mcp.json files

user
disabledMcpjsonServers

MCP servers disabled from .mcp.json files

user
allowedMcpServers

Allowlist of MCP server names that can run

managed
deniedMcpServers

Blocklist of MCP server names that are denied

managed
attribution

Enable attribution comments on generated code

user/project
pr

PR creation defaults (title prefix, label, etc.)

user/project
includeCoAuthoredBy

Add Co-Authored-By line to git commits

user
includeGitInstructions

Include git context instructions in system prompt

user/project
worktree.symlinkPaths

Symlink paths to resolve in git worktree setups

project
hooks

Define pre/post hooks for tool execution events

user/project/managed
disableAllHooks

Globally disable all hook execution

managed
allowManagedHooksOnly

Only allow hooks from managed configuration

managed
allowedHttpHookUrls

Allowlist of URLs for HTTP hooks

managed
httpHookAllowedEnvVars

Env vars that HTTP hooks can access

managed
apiKeyHelper

External command to fetch API key dynamically

user/project/managed
awsCredentialExport

AWS credential export configuration for Bedrock

user
gcpAuthRefresh

GCP auth refresh command for Vertex AI

user
forceLoginMethod

Force a specific login method (oauth, api-key, etc.)

managed
forceLoginOrgUUID

Force login to a specific organization UUID

managed
sshConfigs

SSH configuration for remote sessions

user
companyAnnouncements

Display company-wide announcements in the UI

managed
pluginTrustMessage

Custom trust message shown when enabling plugins

managed
env

Extra environment variables injected into shell sessions

user/project/managed
respectGitignore

Respect .gitignore patterns in file operations

user/project
remote

Remote session configuration (URL, auth, etc.)

user
skipWebFetchPreflight

Skip preflight checks before web fetch operations

user
fileSuggestion

Enable file path suggestions in input

user
$schema

JSON Schema URL for config file validation

user/project
skipAutoPermissionPrompt

Skip the auto-mode permission confirmation prompt

user
useAutoModeDuringPlan

Use auto-mode for planning phase tool calls

user
autoMode.allow

Tools/commands auto-mode can run without asking

user/project/managed
autoMode.soft_deny

Tools auto-mode should avoid but can override

user/project/managed
autoMode.environment

Environment config for auto-mode execution context

user/project/managed
autoUpdatesChannel

Update channel: stable, beta, or nightly

user
minimumVersion

Minimum required version (forces update if older)

managed
disableDeepLinkRegistration

Prevent auto-registration of claude-cli:// protocol

user

06

Environment Variables

70 of the 280 discovered environment variables. Runtime overrides for Claude Code internals.

ANTHROPIC_MODEL

Override the default model

ANTHROPIC_SMALL_FAST_MODEL

Override the small/fast model for quick tasks

ANTHROPIC_BETAS

Comma-separated beta flags to send with API requests

ANTHROPIC_DEFAULT_OPUS_MODEL

Override the default Opus model identifier

ANTHROPIC_DEFAULT_SONNET_MODEL

Override the default Sonnet model identifier

ANTHROPIC_DEFAULT_HAIKU_MODEL

Override the default Haiku model identifier

ANTHROPIC_CUSTOM_MODEL_OPTION

Add a custom model to the model picker

MAX_THINKING_TOKENS

Override max thinking budget

CLAUDE_CODE_MAX_OUTPUT_TOKENS

Override max output tokens

CLAUDE_CODE_EFFORT_LEVEL

Set thinking effort (low/medium/high)

CLAUDE_CODE_SUBAGENT_MODEL

Model for subagent spawning

ANTHROPIC_API_KEY

API key for Anthropic API access

CLAUDE_CODE_OAUTH_TOKEN

OAuth token for Claude Code authentication

CLAUDE_CODE_API_KEY_FILE_DESCRIPTOR

File descriptor to read API key from

CLAUDE_CODE_CUSTOM_OAUTH_URL

Custom OAuth provider URL

ANTHROPIC_AUTH_TOKEN

Alternative auth token for API access

CLAUDE_CODE_DISABLE_ATTACHMENTS

Disable file attachment support

CLAUDE_CODE_DISABLE_CLAUDE_MDS

Disable CLAUDE.md file loading

CLAUDE_CODE_DISABLE_COMMAND_INJECTION_CHECK

Disable command injection safety checks

CLAUDE_CODE_DISABLE_CRON

Disable cron/scheduled task features

07

Hidden CLI Flags

Undocumented command-line flags that unlock experimental modes.

--agent-teams

Enable multi-agent collaboration

--bare

Minimal UI, no chrome

--cowork

Co-working session mode

--spawn

Spawn subagent process

--staging

Use staging API endpoints

--enable-auto-mode

Enable autonomous execution

--chrome-native-host

Chrome native messaging host

--computer-use-mcp

Computer-use MCP server mode

--claude-in-chrome-mcp

Chrome extension MCP bridge

--sdk-url

Override SDK endpoint URL

--fork-session

Fork from existing session

--no-session-persistence

Disable session state persistence

--init-only

Initialize and exit immediately

--deep-link-origin

Deep link source tracking

08

Methodology

Two complementary approaches were used to extract every hidden feature.

AspectNative Binary Analysisnpm Package Analysis
TargetmacOS arm64 Mach-O binary@anthropic-ai/claude-code npm package
Toolsstrings, otool, Ghidra, custom extractorsAST parsing, regex extraction, source maps
String Extraction~42K unique strings from __TEXT segmentFull JS source with readable identifiers
AdvantageIncludes native-only symbols, C++ internalsFull logic, control flow, default values
CoverageSurface-level: names & string constantsDeep: complete implementation details