🔍 Claude Code — Source Code Analysis

Reverse-engineered internals of Anthropic's Claude Code CLI

Unreleased features, secret flags, architecture internals & more

⚠️ Disclaimer: This repository contains AI-assisted analysis of publicly available source code from the @anthropic-ai/claude-code npm package (source maps were accidentally published). Information may be inaccurate, incomplete, or outdated. Feature details and timelines are speculative and may never ship. This is an unofficial community resource and is not affiliated with or endorsed by Anthropic.

Architecture · Tools · Commands · Hidden Features · Secret Flags · Security · Environment Variables

📊 Key Numbers

Metric	Count
Source files	2,000+
Lines of TypeScript	~512K
Built-in tools	43
CLI commands	100+
Service modules	39
React hooks	85
UI components	144
MCP integration files	24
Utility files	564
Permission system files	24

🏗 Architecture

Folder Structure

src/                          2,000+ files · ~512K lines
├── entrypoints/              CLI bootstrap & entry points
│   ├── cli.tsx               Fast-path bootstrap (39KB)
│   ├── init.ts               Initialization sequence
│   ├── mcp.ts                MCP server mode entry
│   └── sdk/                  Agent SDK type definitions
├── commands/                 101 dirs · 207 files · CLI command handlers
│   ├── commit.ts             Git commit command
│   ├── config/               Settings management
│   ├── memory/               Memory CRUD commands
│   └── ...                   (100+ more) doctor, upgrade, login, etc.
├── tools/                    43 tools · 184 files
│   ├── BashTool/             Shell command execution
│   ├── FileReadTool/         Filesystem reading
│   ├── FileEditTool/         Precise string replacement
│   ├── AgentTool/            Subagent spawning & orchestration
│   ├── MCPTool/              Dynamic MCP tool wrapper
│   ├── WebFetchTool/         HTTP fetch + HTML→markdown
│   └── ...                   (37+ more)
├── services/                 39 services · 130 files
│   ├── api/                  Claude API client (Anthropic SDK)
│   ├── mcp/                  MCP protocol (24 files)
│   ├── tools/                StreamingToolExecutor, registry
│   ├── compact/              Context compaction service
│   ├── analytics/            Telemetry & event tracking
│   ├── oauth/                Authentication flows
│   └── plugins/              Community plugin system
├── components/               144 files · Ink terminal UI components
├── hooks/                    85 React hooks (useStream, useTools…)
├── ink/                      Custom terminal renderer
│   ├── reconciler.ts         React reconciler for TTY
│   ├── layout/               Yoga flexbox engine bindings
│   └── termio/               ANSI escape sequence parser
├── utils/                    564 files · Helper functions
│   ├── permissions/          Permission system (24 files)
│   ├── bash/                 Bash AST parser (safety checks)
│   ├── model/                Model selection & routing
│   ├── settings/             Settings read/write layer
│   └── git.ts                Git porcelain operations
├── state/                    Centralized application state
├── bridge/                   Remote control protocol (31 files, ~400KB)
├── cli/                      Transport layer (WS/SSE)
├── memdir/                   Memory persistence layer
├── tasks/                    Background task system
├── coordinator/              Multi-agent orchestration
├── assistant/                KAIROS persistent mode
├── buddy/                    AI Companion Pet system
├── voice/                    Voice input pipeline
├── vim/                      Full Vim emulation
└── schemas/                  JSON validation schemas

Boot Sequence

What happens when you type claude:

1. cli.tsx loads         → Fast-path: --version exits with zero imports (<50ms)
2. Feature flags         → bun:bundle eliminates dead code at build time
3. main.tsx initializes  → Commander.js parses argv, routes to handler
4. Config loaded         → settings.json + CLAUDE.md files + .envrc via direnv
5. Auth checked          → OAuth token or ANTHROPIC_API_KEY validated
6. GrowthBook init       → Remote feature flags fetched async
7. Tools assembled       → 43 built-in tools loaded from registry
8. MCP servers connected → stdio/SSE/WebSocket negotiated in parallel
9. System prompt built   → 10+ component sources assembled
10. REPL launched        → Ink renders terminal UI via React reconciler + Yoga
11. Query loop begins    → Event loop idles, background tasks on timers

Query Loop — The Core Cycle

User types message → createUserMessage()
    ↓
Append to conversation history (in-memory array)
    ↓
Build system prompt (CLAUDE.md + tools + context + memory)
    ↓
Stream to Claude API (Anthropic SDK · SSE)
    ↓
Parse response tokens (incremental terminal render)
    ↓
if tool_use blocks found → findToolByName() → canUseTool()
    ↓
StreamingToolExecutor (concurrent-safe, parallel-capable)
    ↓
Collect tool results → loop back to API
    ↓
Display final response (markdown rendered)
    ↓
Post-sampling hooks (auto-compact · memory extract · dream)
    ↓
Wait for next input (cycle restarts)

Key implementation details:

Streaming by default — tokens render as they arrive
Parallel tool execution — StreamingToolExecutor runs concurrent tools
Agentic loop — cycles until Claude emits no tool_use blocks
React rendering — Ink reconciler + Yoga flexbox, ANSI diffs minimize redraws
Interrupt handling — Ctrl+C aborts executor, preserves conversation
Post-sampling hooks — compaction check, memory extraction, dream consolidation

🔧 Built-in Tools (43)

File Operations & Execution

Tool	Description	Category
`FileReadTool`	Read files from filesystem	Core
`FileEditTool`	Precise string replacement edits	Core
`FileWriteTool`	Write new files	Core
`GlobTool`	File pattern matching	Core
`GrepTool`	Content search (ripgrep-backed)	Core
`NotebookEditTool`	Jupyter notebook editing	Core
`BashTool`	Shell command execution	Core
`PowerShellTool`	Windows PowerShell execution	Platform

Agents, Planning & MCP

Tool	Description	Category
`AgentTool`	Subagent spawning & orchestration	Core
`SendMessageTool`	Inter-agent communication	Multi-agent
`TaskCreateTool`	Create background tasks	Task
`TaskGetTool`	Retrieve task status	Task
`TaskListTool`	List all tasks	Task
`TaskUpdateTool`	Update task progress	Task
`TaskStopTool`	Stop running tasks	Task
`TaskOutputTool`	Get task output	Task
`EnterPlanModeTool`	Enter planning phase	Planning
`ExitPlanModeV2Tool`	Exit planning phase	Planning
`EnterWorktreeTool`	Create git worktree	Worktree
`ExitWorktreeTool`	Exit git worktree	Worktree
`MCPTool`	Dynamic MCP tool wrapper	MCP
`ListMcpResourcesTool`	List MCP resources	MCP
`ReadMcpResourceTool`	Read MCP resources	MCP

Search & Fetch

Tool	Description	Category
`WebFetchTool`	HTTP fetch + HTML→markdown	Core
`WebSearchTool`	Web search	Core
`ToolSearchTool`	Search available tools	Feature-gated

System & Utility

Tool	Description	Category
`AskUserQuestionTool`	Prompt user for input	Core
`TodoWriteTool`	Write TODO files	Core
`SkillTool`	Invoke skills	Core
`BriefTool`	Brief/summary output	Core
`ConfigTool`	Settings management	ANT-only
`TungstenTool`	Virtual terminal	ANT-only

Feature-Gated / Experimental

These tools are compiled into the binary but only surface when the corresponding feature flag is enabled:

Tool	Gate	Description
`SleepTool`	`PROACTIVE` / `KAIROS`	Pause between actions
`WebBrowserTool`	`WEB_BROWSER_TOOL`	Browser automation
`TerminalCaptureTool`	`TERMINAL_PANEL`	Terminal screen capture
`CtxInspectTool`	`CONTEXT_COLLAPSE`	Context inspection
`SnipTool`	`HISTORY_SNIP`	Snippet compression
`WorkflowTool`	`WORKFLOW_SCRIPTS`	Workflow automation
`ListPeersTool`	`UDS_INBOX`	List peer sessions
`OverflowTestTool`	`OVERFLOW_TEST_TOOL`	Overflow testing
`MonitorTool`	`MONITOR_TOOL`	Monitoring
`VerifyPlanExecutionTool`	`CLAUDE_CODE_VERIFY_PLAN`	Plan verification
`SuggestBackgroundPRTool`	`USER_TYPE=ant`	Suggest background PRs
`REPLTool`	`USER_TYPE=ant`	REPL mode

⌨️ Slash Commands

Publicly Available Commands

Command	Description
`/help`	Show help text
`/compact`	Compact conversation context
`/clear`	Clear conversation
`/config`	Settings management
`/cost`	Show session cost
`/doctor`	Diagnostic checks
`/init`	Initialize project
`/login` / `/logout`	Authentication
`/memory`	Memory management
`/model`	Change model
`/mcp`	MCP server management
`/resume`	Resume session
`/vim`	Toggle vim mode
`/theme`	Change theme
`/plan`	Toggle plan mode
`/permissions`	Manage permissions
`/diff`	Show file changes
`/review`	Code review
`/export`	Export transcript
`/upgrade`	Upgrade CLI
`/status`	Show status
`/hooks`	Manage hooks
`/plugin`	Plugin management
`/skills`	Skills management
`/agents`	Agent management
`/fast`	Toggle fast mode

Hidden / Internal Slash Commands (26)

Not shown in --help, discovered via source analysis

Command	Description	Access
`/ctx-viz`	Context visualizer	Internal
`/btw`	Side question (quick note)	Available
`/good-claude`	Easter egg	Internal
`/teleport`	Session teleport	Internal
`/share`	Session sharing	Internal
`/summary`	Summarize conversation	Internal
`/ultraplan`	Advanced 30-min planning	Feature-gated
`/subscribe-pr`	PR webhooks	Feature-gated
`/autofix-pr`	Auto PR fix	Internal
`/ant-trace`	Internal trace	ANT-only
`/perf-issue`	Performance report	Internal
`/debug-tool-call`	Debug tool calls	Internal
`/bughunter`	Bug debugging	Internal
`/force-snip`	Force snippet	Feature-gated
`/mock-limits`	Mock rate limits	Internal
`/bridge-kick`	Bridge test	Internal
`/backfill-sessions`	Backfill sessions	Internal
`/break-cache`	Invalidate cache	Internal
`/agents-platform`	Agent platform	ANT-only
`/onboarding`	Setup flow	Internal
`/oauth-refresh`	Token refresh	Internal
`/env`	Environment inspect	Internal
`/reset-limits`	Reset rate limits	Internal
`/dream`	Memory consolidation	Feature-gated
`/version`	Internal version	Internal
`/init-verifiers`	Verifier setup	Internal

🔮 Unreleased Features

🐾 BUDDY — AI Companion Pet

Easter Egg · Unreleased

Every user gets a unique virtual pet generated deterministically from their account ID. The buddy system includes species, rarity tiers, stats, cosmetics, and animations.

Verified in source: src/buddy/types.ts, src/buddy/companion.ts, src/buddy/CompanionSprite.tsx

Property	Details
Release	Teaser: Apr 1–7, 2026 · Live: May 2026
Species	duck, goose, blob, cat, dragon, octopus, owl, penguin, turtle, snail, ghost, axolotl, capybara, cactus, robot, rabbit, mushroom, chonk
Rarity	Common (60%), Uncommon (25%), Rare (10%), Epic (4%), Legendary (1%)
Eyes	`· ✦ × ◉ @ °`
Hats	none, crown, tophat, propeller, halo, wizard, beanie, tinyduck
Stats	DEBUGGING, PATIENCE, CHAOS, WISDOM, SNARK
Soul	Generated once per user, stored forever in config
Shiny	~1% chance of shiny variant

Fun fact: The species names are encoded as String.fromCharCode() calls in the source to evade build-time string filters (e.g. "capybara" collides with a model codename).

🔮 KAIROS — Persistent Assistant

Unreleased · Feature gate: feature('KAIROS') + tengu_kairos

An always-on mode where Claude remembers everything across sessions with daily logs and overnight "dream" memory consolidation.

Property	Details
Daily Logs	`~/.claude/.../logs/YYYY/MM/DD.md`
Log mode	append-only
Dream bash	Read-only
Blocking budget	15s max — auto-backgrounds
Status modes	normal · proactive
Exclusive Tools	`SendUserFile`, `PushNotification`, `SubscribePR`, `SleepTool`

Dream Phases: Orient → Gather → Consolidate → Prune

⚡ ULTRAPLAN — 30-Min Remote Planning

Unreleased · src/commands/ultraplan.tsx (66KB)

For complex tasks, Claude spins up a separate cloud instance that explores and plans for up to 30 minutes. Users review and approve the plan in their browser.

Property	Details
Model	Opus 4.6 via `tengu_ultraplan_model`
Poll interval	3s
Flow	Poll → ExitPlanMode → approve/reject → loop or execute
Teleport	"Teleport to terminal" archives remote, runs locally

⚙️ Coordinator Mode — Multi-Agent

ENV activated: CLAUDE_CODE_COORDINATOR_MODE=1

Claude becomes a manager, breaking tasks into pieces and assigning each to parallel worker agents.

Verified in source: src/coordinator/, src/tools.ts lines 120-134

Property	Details
Protocol	`<task-notification>` XML
Isolation	Scratch dirs via `tengu_scratch`
Continue	via `SendMessage`
Notification fields	status, summary, tokens, duration

📬 UDS Inbox — Cross-Session IPC

Unreleased · Feature gate: UDS_INBOX

Multiple Claude sessions on the same machine can send messages to each other like a team chat.

Property	Details
Teammate	`to: "researcher"`
Local socket	`to: "uds:/.../sock"`
Remote	`to: "bridge:..."`
Discovery	`ListPeersTool` reads `~/.claude/sessions/`

📱 Bridge — Remote Control

Unreleased · Feature gate: BRIDGE_MODE

Run Claude locally but control it from your phone or claude.ai in the browser.

Verified in source: src/bridge/ (31 files, ~400KB total)

Property	Details
Command	`claude remote-control`
Init API	`POST /v1/environments/bridge`
Transport	poll → WebSocket
Control messages	`initialize`, `set_model`, `can_use_tool`
JWT refresh	Every 3h55m (tokens last ~4hrs)

👾 Daemon Mode — Session Supervisor

Unreleased · Feature gate: DAEMON

Run Claude sessions in the background like system services.

Property	Details
Background	`claude --bg <prompt>` in tmux
On exit	detach (session persists)
Commands	`daemon`, `ps`, `logs`, `attach`, `kill`

🌙 Auto-Dream — Memory Consolidation

Unreleased · Feature gate: tengu_onyx_plover

Between sessions, Claude reviews what it learned and organizes notes into structured memory files.

Property	Details
Trigger	≥24h + ≥5 sessions since last dream
Output limit	<25KB
Phases	Orient → Gather → Consolidate → Prune

🚩 Secret CLI Flags

Flag	Description	Status
`--bare`	No hooks/plugins/memory	Available
`--dump-system-prompt`	Print system prompt & exit	Available
`--daemon-worker=<k>`	Daemon subprocess	Unreleased
`--computer-use-mcp`	Computer Use MCP server	Unreleased
`--claude-in-chrome-mcp`	Chrome MCP	Unreleased
`--chrome-native-host`	Chrome native messaging	Unreleased
`--bg`	Background tmux session	Available
`--spawn`	Spawn mode	Unreleased
`--capacity <n>`	Parallel worker limit	Unreleased
`--worktree / -w`	Git worktree isolation	Available

🔬 Build-Time Feature Flags

32 feature flags compiled into the binary via bun:bundle. Dead code elimination removes disabled features from external builds.

import { feature } from "bun:bundle";

// Example: conditionally load a tool
const SleepTool =
  feature("PROACTIVE") || feature("KAIROS")
    ? require("./tools/SleepTool/SleepTool.js").SleepTool
    : null;

Flag	Description
`KAIROS`	Persistent assistant mode
`PROACTIVE`	Background sleeping agents
`COORDINATOR_MODE`	Multi-agent orchestration
`BRIDGE_MODE`	Remote control
`DAEMON`	Session supervisor
`BG_SESSIONS`	Background sessions
`ULTRAPLAN`	30-min cloud planning
`BUDDY`	AI companion pet
`TORCH`	??? (opaque)
`WORKFLOW_SCRIPTS`	Workflow automation
`VOICE_MODE`	Voice input · `/voice` works
`TEMPLATES`	Job templates
`CHICAGO_MCP`	Computer use (live for Max/Pro)
`UDS_INBOX`	Unix socket IPC
`REACTIVE_COMPACT`	Realtime compaction
`CONTEXT_COLLAPSE`	Smart context collapse
`HISTORY_SNIP`	Snippet compression
`CACHED_MICROCOMPACT`	Cached micro-compaction
`TOKEN_BUDGET`	Per-turn token budgets
`EXTRACT_MEMORIES`	Background memory extraction
`OVERFLOW_TEST_TOOL`	Overflow testing
`TERMINAL_PANEL`	Terminal screen capture
`WEB_BROWSER_TOOL`	Browser automation
`FORK_SUBAGENT`	Agent forking
`DUMP_SYS_PROMPT`	Print system prompt
`ABLATION_BASE`	Research mode
`BYOC_RUNNER`	Bring-your-own-compute runner
`SELF_HOSTED`	Self-hosted mode
`MONITOR_TOOL`	Monitoring
`CCR_AUTO`	Auto cloud compute
`MEM_SHAPE_TEL`	Memory analytics
`SKILL_SEARCH`	Experimental skill search

🌍 Environment Variables

🐛 Debug & Profiling

Variable	Description
`CLAUDE_CODE_PERFETTO_TRACE`	Chrome trace via Perfetto
`CLAUDE_CODE_PROFILE_STARTUP`	Startup timing profiler
`CLAUDE_CODE_FRAME_TIMING_LOG`	Frame timing log output
`CLAUDE_CODE_VCR_RECORD`	Record HTTP interactions
`CLAUDE_CODE_DEBUG_REPAINTS`	Visualize UI repaints

🎛️ Runtime Overrides

Variable	Description
`CLAUDE_CODE_OVERRIDE_DATE`	Inject fake date
`CLAUDE_CODE_MAX_CONTEXT_TOKENS`	Override context window
`MAX_THINKING_TOKENS`	Override thinking budget
`CLAUDE_CODE_EXTRA_BODY`	Inject extra API params
`AUTOCOMPACT_PCT_OVERRIDE`	Override compact threshold
`IDLE_THRESHOLD_MINUTES`	Idle threshold (75m default)

⚠️ Safety Bypass (Dangerous)

Variable	Description
`DISABLE_COMMAND_INJECTION_CHECK`	Skip injection guard — DANGEROUS
`CLAUDE_CODE_ABLATION_BASELINE`	Disable ALL safety features
`DISABLE_INTERLEAVED_THINKING`	Disable interleaved thinking

🔑 Anthropic Internal

Variable	Description
`USER_TYPE=ant`	Unlock all internal features
`CLAUDE_INTERNAL_FC_OVERRIDES`	Override feature flags
`CLAUDE_MORERIGHT`	"More right" layout
`CLAUDE_CODE_UNDERCOVER`	Undercover mode
`CLAUBBIT`	Internal testing

🔒 Security & Permissions

3-Layer Permission Model

Layer 1: Tool Registry Filter
  └── filterToolsByDenyRules() removes denied tools BEFORE Claude sees them
  └── Claude cannot call tools blocked at this layer

Layer 2: Per-call Permission Check
  └── canUseTool() checks allow/deny rules per invocation
  └── Checks tool name, arguments, and working directory

Layer 3: Interactive User Prompt
  └── If no rule matches → user prompted
  └── Options: allow once · allow always · deny

Permission Modes:

default — interactive prompting
plan — requires explicit approval
auto — AI classifier decides
bypassPermissions — full bypass
dontAsk — auto-deny
acceptEdits — auto-accept file edits

Bash Safety — AST-Level Analysis

The Bash tool includes a full shell AST parser in utils/bash/. Before executing any command, it parses the AST to detect dangerous patterns:

rm -rf / — destructive deletion
Fork bombs
curl | bash — remote code execution
sudo escalation
TTY injection
History manipulation
sleep N where N >= 2s is blocked

Prompt Injection Defenses (6 Layers)

System prompt warning — instructs model to flag suspected injection
Unicode smuggling prevention — partiallySanitizeUnicode() strips hidden chars
Subprocess env protection — prevents secret exfiltration via untrusted content
Tool result budgeting — toolResultStorage.ts prevents overflow attacks
Trusted/untrusted separation — hooks = trusted, tool results = potentially untrusted
Hook trust gates — ALL hooks require workspace trust dialog acceptance

Settings.json Rule Format

{
  "permissions": {
    "allow": ["Bash(git *)"],
    "deny": ["Bash(rm *)", "Write"]
  }
}

🧠 Models & Intelligence

Complete Model Registry

Family	Versions	Providers
Opus	4.6, 4.5, 4.1, 4.0	1P, Bedrock, Vertex, Foundry
Sonnet	4.6, 4.5, 4.0, 3.7, 3.5	1P, Bedrock, Vertex, Foundry
Haiku	4.5, 3.5	1P, Bedrock, Vertex, Foundry

Model aliases: opus, sonnet, haiku, best (Opus 4.6), opusplan (Opus in plan mode else Sonnet)

Model Selection Priority (5 layers)

1. /model session override     (highest)
2. --model flag
3. ANTHROPIC_MODEL env
4. settings file
5. built-in default            (lowest)

Max/Team subscribers → Opus 4.6
Standard users → Sonnet 4.6
Anthropic employees (Ants) → Opus 4.6 [1m]

Fast Mode

Verified in source: src/utils/fastMode.ts

Property	Details
Supported model	Opus 4.6 only
Beta header	`fast-mode-2026-02-01`
Pricing	$30/$150 per MTok vs $5/$25 normal (6× markup)
Availability	1P only (not Bedrock/Vertex/Foundry)
Disable	`CLAUDE_CODE_DISABLE_FAST_MODE=1`
Default	Enabled for paying users

UltraThink Enhanced Reasoning

Verified in source: src/utils/thinking.ts

Property	Details
Modes	adaptive (model decides), enabled (fixed budget), disabled
Models	Opus 4.6 and Sonnet 4.6 support adaptive thinking
Trigger	`hasUltrathinkKeyword()` detects "ultrathink" in user input
Visual	7-color rainbow shimmer animation
Gate	Compile-time `feature('ULTRATHINK')` + runtime `tengu_turtle_carbon`

Context Window & Tokens

Property	Value
Default context	200,000 tokens
Extended context	1M via `[1m]` model suffix
1M gate	Beta header `context-1m-2025-08-07`
Opus 4.6 output	64K default, 128K upper limit
Sonnet 4.6 output	32K default, 128K upper limit
Other models	32K default, 64K upper limit
HIPAA disable	`CLAUDE_CODE_DISABLE_1M_CONTEXT`

AutoCompact & Context Management

Verified in source: src/services/compact/autoCompact.ts

Property	Value
Buffer tokens	13,000
Trigger	~187K of 200K (~80% usage)
Circuit breaker	3 consecutive failures
MicroCompact	Bash, FileRead, Grep output only
Post-compact	Restores top-5 files (5K tokens each, 25K total)
Warning threshold	20,000 buffer tokens
Manual compact	3,000 buffer tokens

🌱 GrowthBook Feature Gates

Gradual rollout gates in the tengu_* namespace:

Gate	Function
`tengu_malort_pedway`	Computer use
`tengu_onyx_plover`	Auto-dream
`tengu_kairos`	Assistant mode
`tengu_ultraplan_model`	Planning model
`tengu_cobalt_raccoon`	Auto-compact
`tengu_portal_quail`	Memory extract
`tengu_harbor`	MCP allowlist
`tengu_scratch`	Worker scratch dirs
`tengu_herring_clock`	Team memory
`tengu_chomp_inflection`	Prompt suggestions
`tengu_turtle_carbon`	UltraThink
`tengu_pewter_ledger`	Plan file A/B test
`tengu_marble_sandcastle`	Fast mode binary requirement

🧪 API Beta Headers

anthropic-beta header values used by Claude Code:

Feature	Header Date
Interleaved Thinking	`2025-05-14`
1M Context Window	`2025-08-07`
Structured Outputs	`2025-12-15`
Advanced Tool Use	`2025-11-20`
Tool Search	`2025-10-19`
Effort Levels	`2025-11-24`
Task Budgets	`2026-03-13`
Fast Mode	`2026-02-01`
Prompt Cache Scoping	`2026-01-05`
CLI Internal (ant)	`2026-02-09`

🔌 Extension Points

MCP Integration

24 files, ~500KB of MCP client code in src/services/mcp/

Property	Details
Transports	stdio, SSE, HTTP, WebSocket, SDK, CloudAI proxy
Server scopes	local, user, project, dynamic, enterprise, claudeai, managed
Auth	OAuth with callback ports, XAA, token refresh
Tool naming	`mcp__<server>__<tool>`
Connection states	connected, failed, needs-auth, pending, disabled

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/me/projects"
      ]
    },
    "remote-api": {
      "url": "https://api.example.com/mcp",
      "transport": "sse"
    }
  }
}

Hooks System (20 Events)

Verified in source: src/utils/hooks.ts (5,023 lines, 159KB)

Event	Description
`Setup`	Initial setup
`SessionStart` / `SessionEnd`	Session lifecycle
`PreToolUse` / `PostToolUse`	Before/after tool execution
`PostToolUseFailure`	After tool failure
`StopHook`	Stop condition
`PreCompact` / `PostCompact`	Before/after compaction
`SubagentStart` / `SubagentStop`	Subagent lifecycle
`TeammateIdle`	Teammate idle notification
`TaskCreated` / `TaskCompleted`	Task lifecycle
`ConfigChange`	Configuration changed
`CwdChanged`	Working directory changed
`FileChanged`	File modification
`UserPromptSubmit`	User submits prompt
`PermissionRequest` / `PermissionDenied`	Permission events

Tool hook timeout: 10 minutes
SessionEnd timeout: 1.5s (overridable via env)
Trust: ALL hooks require workspace trust dialog
Async: asyncRewake hooks enqueue task-notification on completion
Exit code 2 = blocking error

CLAUDE.md Loading Hierarchy

Priority (low → high):
1. /etc/claude-code/CLAUDE.md          (system-wide)
2. ~/.claude/CLAUDE.md                 (user-level)
3. project/CLAUDE.md                   (project-level)
   project/.claude/CLAUDE.md
   project/.claude/rules/*.md
4. CLAUDE.local.md                     (local override, highest)

@include directive: @path, @./relative, @~/home, @/absolute
Frontmatter paths: field for glob-based applicability
HTML comment stripping for  blocks
Circular reference prevention via tracking

Skills System

Type	Description
Bundled	~19 compiled-in skills (batch, claudeApi, debug, keybindings, loop, remember, simplify, skillify, etc.)
Disk-based	Loaded from `.claude/skills/` directories
MCP-based	Skills from MCP servers

Plugin System

BUILTIN_PLUGINS map with per-plugin enable/disable
Each plugin provides: Skills, Hooks, MCP servers
Source tagging: builtin@builtin vs marketplace@npm
User toggle via settings.enabledPlugins[pluginId]

🏗️ Infrastructure Internals

Multi-Agent Swarm System

Backend	Description
InProcessBackend	Same Node.js process via `AsyncLocalStorage` (zero overhead)
TmuxBackend	Separate tmux panes/windows
ITermBackend	iTerm2 split panes

Teams: .claude/teams/<name>.json with lead/member config
Identity: agentName@teamName
Tool restrictions: agents cannot use TaskOutput, ExitPlanMode, AskUserQuestion, Workflow

Memory System (4 Types)

Type	Scope	Default
`user`	Always private	Private
`feedback`	Default private	Private
`project`	Bias toward team	Team
`reference`	Usually team	Team

Storage format: YAML frontmatter + Markdown content

Vim Mode

Full state machine with INSERT and NORMAL modes:

Operators: d (delete), c (change), y (yank), >/< (indent)
Motions: h/j/k/l, w/b/e, 0/^/$
Text objects: 11 types — w, W, ", ', (, ), {, }, [, ], <, >
Find: f/F/t/T with repeat
Dot-repeat with full change recording
Max count: 10,000

Cost Tracking

Verified in source: src/cost-tracker.ts

Tracks per-session:

Input/output/cache tokens per model
Web search request count
USD cost via calculateUSDCost() with per-model pricing
Lines of code added/removed
API duration vs wall clock duration
Stored in .claude/config.json indexed by session ID

Internal Codename Map

Codename	Meaning
Capybara	Active internal model (encoded as charCodes to evade filters)
Fennec	Retired Opus codename
Turtle Carbon	UltraThink feature gate
Pewter Ledger	Plan file structure A/B test
Birch Mist	First-message optimization experiment
Penguin Mode	Fast mode (internal name)

🗝️ Hardcoded SDK API Keys

These SDK keys are baked into the binary for specific Anthropic environments. They are not user credentials.

Environment	Key Prefix
Ant prod	`sdk-xRVcrliHIlrg4og4`
Ant dev	`sdk-yZQvlplybuXjYh6L`
External	`sdk-zAZezfDKGoZuXXKe`

📚 Source References

Key source files for deeper exploration:

File	Description	Size
`src/main.tsx`	Main application	803KB
`src/tools.ts`	Tool registry	17KB
`src/commands.ts`	Command registry	25KB
`src/Tool.ts`	Tool interface (793 lines)	29KB
`src/QueryEngine.ts`	Query engine	46KB
`src/utils/hooks.ts`	Hooks system	159KB
`src/utils/messages.ts`	Message handling	193KB
`src/utils/sessionStorage.ts`	Session storage	180KB
`src/bridge/bridgeMain.ts`	Bridge orchestrator	115KB
`src/bridge/replBridge.ts`	Bridge REPL	100KB
`src/utils/teleport.tsx`	Teleport system	175KB
`src/utils/attachments.ts`	Attachments	127KB
`src/utils/ansiToPng.ts`	ANSI to PNG export	214KB
`src/commands/ultraplan.tsx`	UltraPlan impl	66KB
`src/commands/insights.ts`	Usage insights	115KB
`src/utils/fastMode.ts`	Fast mode impl	17KB
`src/utils/thinking.ts`	UltraThink	5KB
`src/cost-tracker.ts`	Cost tracking	10KB
`src/buddy/types.ts`	Buddy pet types	3KB
`src/buddy/CompanionSprite.tsx`	Pet rendering	45KB

🙏 Credits

Original source: @anthropic-ai/claude-code npm package (source maps accidentally published)
Analysis inspiration: ccleaks.com by Abhishek Tiwari (@baanditeagle)
Initial discovery: @Fried_rice on X

⚖️ Legal Notice

This repository is for educational and research purposes only. All code and analysis is based on publicly available source code that was accidentally published by Anthropic via npm source maps. No proprietary systems were accessed. Feature details are speculative and may never be released.

If you are from Anthropic and would like this repository taken down, please open an issue or contact the repository owner.

⭐ Star this repo if you found it useful!

Last updated: March 31, 2026

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
docs		docs
src		src
.gitignore		.gitignore
README.md		README.md

Folders and files

Latest commit

History

Repository files navigation

🔍 Claude Code — Source Code Analysis

📊 Key Numbers

🏗 Architecture

Folder Structure

Boot Sequence

Query Loop — The Core Cycle

🔧 Built-in Tools (43)

File Operations & Execution

Agents, Planning & MCP

Search & Fetch

System & Utility

Feature-Gated / Experimental

⌨️ Slash Commands

Publicly Available Commands

Hidden / Internal Slash Commands (26)

🔮 Unreleased Features

🐾 BUDDY — AI Companion Pet

🔮 KAIROS — Persistent Assistant

⚡ ULTRAPLAN — 30-Min Remote Planning

⚙️ Coordinator Mode — Multi-Agent

📬 UDS Inbox — Cross-Session IPC

📱 Bridge — Remote Control

👾 Daemon Mode — Session Supervisor

🌙 Auto-Dream — Memory Consolidation

🚩 Secret CLI Flags

🔬 Build-Time Feature Flags

🌍 Environment Variables

🐛 Debug & Profiling

🎛️ Runtime Overrides

⚠️ Safety Bypass (Dangerous)

🔑 Anthropic Internal

🔒 Security & Permissions

3-Layer Permission Model

Bash Safety — AST-Level Analysis

Prompt Injection Defenses (6 Layers)

Settings.json Rule Format

🧠 Models & Intelligence

Complete Model Registry

Model Selection Priority (5 layers)

Fast Mode

UltraThink Enhanced Reasoning

Context Window & Tokens

AutoCompact & Context Management

🌱 GrowthBook Feature Gates

🧪 API Beta Headers

🔌 Extension Points

MCP Integration

Hooks System (20 Events)

CLAUDE.md Loading Hierarchy

Skills System

Plugin System

🏗️ Infrastructure Internals

Multi-Agent Swarm System

Memory System (4 Types)

Vim Mode

Cost Tracking

Internal Codename Map

🗝️ Hardcoded SDK API Keys

📚 Source References

🙏 Credits

⚖️ Legal Notice

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages