Claude Code: 330+ Environment Variables and 32 Feature Flags Explained

Most developers use Claude Code with default settings. The leaked source reveals they are barely scratching the surface. 330+ environment variables control everything from model selection to stealth mode. 32 compile-time feature flags gate unreleased systems like KAIROS, BUDDY, and VOICE. 22+ GrowthBook runtime gates enable A/B testing of features for specific user segments. This is the complete configuration reference — every variable, every flag, every hidden setting discovered in the v2.1.88 source leak.

The leak originated from the @anthropic-ai/claude-code npm package, which shipped a 59.8 MB sourcemap file (cli.js.map) containing the full reconstructable TypeScript source. Security researcher Chaofan Shou identified the exposure, revealing approximately 1,900 source files, 43 built-in tools, 26 hidden slash commands, and the internal codename for the project: Tengu (a supernatural creature from Japanese mythology).

32 Compile-Time Feature Flags

Claude Code uses Bun's feature() function at bundle time. These flags trigger constant-folding and dead-code elimination — in external builds, disabled flags compile to false and associated code is stripped entirely. This means the unreleased features exist in the source but never execute in the production binary.

Additional Flags Identified in Extended Analysis

Beyond the core 32 flags, several additional compile-time flags were identified:

• NATIVE_CLIENT_ATTESTATION — Native client attestation (placeholder hash cch=00000)
• ANTI_DISTILLATION_CC — Injects fake decoy tools into API requests to poison training data of competing models
• TRANSCRIPT_CLASSIFIER — Automatic session transcript classification
• KAIROS_BRIEF — Brief mode for the persistent KAIROS assistant
• KAIROS_GITHUB_WEBHOOKS — GitHub integration via SubscribePRTool for autonomous PR reactions
• LODESTONE — Undocumented feature
• CLAUDE_CODE_VERIFY_PLAN — Plan execution verification tool

22+ Runtime Feature Gates via GrowthBook

While compile-time flags are baked into the binary, runtime feature gates use GrowthBook for remote control. All gates are prefixed with tengu_ (the project's internal codename) and many use obfuscated random word pairs like tengu_frond_boric to hide their purpose. Values are aggressively cached via getFeatureValue_CACHED_MAY_BE_STALE() to avoid blocking the main loop.

Critical implication: Anthropic can enable or disable these flags remotely without requiring a user update, via hourly polling to a remote settings endpoint.

Remote Kill-Switches

The codebase contains 6+ remote kill-switches capable of:

• Bypassing permission prompts
• Enabling/disabling Fast Mode globally
• Toggling Voice Mode
• Controlling analytics collection
• Forcing complete application exit

330+ Environment Variables: Complete Reference by Category

The environment variables fall into 31 distinct categories. Below we cover the most significant ones that developers and power users can actually leverage.

Authentication and API (25 variables)

These control how Claude Code authenticates with the Anthropic API, including API keys, OAuth tokens, and custom authentication workflows.

Model Configuration (18 variables)

The model configuration layer is where power users get the most leverage. These variables control which models are used for different tasks.

Cloud Providers (33 variables)

Claude Code supports three alternative cloud providers beyond the direct Anthropic API: AWS Bedrock (16 vars), Google Vertex AI (12 vars), and Microsoft Azure Foundry (5 vars).

export CLAUDE_CODE_USE_BEDROCK=1
export AWS_REGION=us-east-1
export AWS_ACCESS_KEY_ID=your_key
export AWS_SECRET_ACCESS_KEY=your_secret
# Optional: 1-hour prompt caching
export ENABLE_PROMPT_CACHING_1H_BEDROCK=1

export CLAUDE_CODE_USE_VERTEX=1
export ANTHROPIC_VERTEX_PROJECT_ID=your_project_id
export CLOUD_ML_REGION=us-central1

export CLAUDE_CODE_USE_FOUNDRY=1
export ANTHROPIC_FOUNDRY_BASE_URL=https://your-endpoint.azure.com
export ANTHROPIC_FOUNDRY_API_KEY=your_key

Behavior and UX (20+ variables)

Debug and Profiling (8 variables)

Tools and Limits (12 variables)

Context and Compaction (6 variables)

Agent Teams and Multi-Agent (8 variables)

These variables control Claude Code's multi-agent orchestration system, where a lead agent coordinates parallel workers.

Stealth and Privacy (5 variables)

export CLAUDE_CODE_UNDERCOVER=1

OpenTelemetry Observability (22 variables)

Claude Code ships with full OpenTelemetry support, allowing developers to send traces, metrics, and logs to their own monitoring stack (Grafana, Datadog, Jaeger, etc.).

# Send everything to your own OTEL stack
export OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317
export OTEL_TRACES_EXPORTER=otlp
export OTEL_METRICS_EXPORTER=otlp
export OTEL_LOGS_EXPORTER=otlp

# Full visibility into tool usage and prompts
export OTEL_LOG_TOOL_DETAILS=1
export OTEL_LOG_TOOL_CONTENT=1
export OTEL_LOG_USER_PROMPTS=1

Proxy and TLS (10 variables)

Employee-Only: USER_TYPE=ant

The USER_TYPE === 'ant' guard restricts specific features to Anthropic employees. When active, employees gain access to:

15 API Beta Headers

These headers are sent in the anthropic-beta field of API requests. They follow the format feature-name-YYYY-MM-DD.

Anti-Distillation in API Requests

When the ANTI_DISTILLATION_CC flag is active, Claude Code sends an anti_distillation field with the value ["fake_tools"]. This instructs the server to silently inject fake tool definitions into the system prompt. The purpose: if someone captures the API traffic to train a competing model, the fake tool definitions poison their training data.

The Most Useful Variables You Can Set Right Now

For developers who want to optimize their Claude Code experience immediately, here is the recommended power-user shell profile:

# Model: Force Opus for main, Sonnet for sub-agents (cost savings)
export ANTHROPIC_MODEL=claude-opus-4-6
export CLAUDE_CODE_SUBAGENT_MODEL=claude-sonnet-4-6
export FALLBACK_FOR_ALL_PRIMARY_MODELS=claude-sonnet-4-6

# Performance: Max thinking, max output
export MAX_THINKING_TOKENS=32000
export CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000
export CLAUDE_CODE_EFFORT_LEVEL=high

# Agent Teams: Enable multi-agent with 3 workers
export CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
export CLAUDE_CODE_PLAN_V2_AGENT_COUNT=3

# Privacy: Zero telemetry, zero surveys
export DISABLE_TELEMETRY=1
export DISABLE_AUTOUPDATER=1
export CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY=1
export CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1

# Tools: Higher concurrency and limits
export CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY=5
export BASH_MAX_OUTPUT_LENGTH=50000
export MCP_TIMEOUT=60000
export MCP_TOOL_TIMEOUT=120000

# Debug (enable when needed)
# export CLAUDE_DEBUG=1
# export CLAUDE_CODE_PERFETTO_TRACE=1

Complete Variable Count by Category

Frequently Asked Questions