Claude Code Feature Flags and Environment Variables¶

Claude Code exposes environment variables that control model selection, context handling, tool concurrency, and observability — each prefixed with a naming convention (ENABLE_*, DISABLE_*, CLAUDE_CODE_*) that signals its stability and risk profile.

The Three-Tier Naming Convention¶

Claude Code uses a naming pattern that signals each variable's risk profile (settings docs):

Prefix	Meaning	Signal
`ENABLE_*`	Opt-in to experiments	Not default — expect rough edges
`DISABLE_*`	Opt-out of defaults	Default, suppressible
`CLAUDE_CODE_*`	Behavior parameters	Stable — safe to tune

The pattern mirrors progressive-exposure feature flagging: new capabilities ship behind ENABLE_* before flipping to default.

Performance and Model Control¶

Variable	Effect	Default
`CLAUDE_CODE_EFFORT_LEVEL`	Adaptive reasoning level: `low`, `medium`, `high`	`auto` (env-vars docs)
`MAX_THINKING_TOKENS`	Extended-thinking budget (used when adaptive is off)	Model-dependent
`CLAUDE_CODE_MAX_OUTPUT_TOKENS`	Cap on output tokens per response	Model default (env-vars docs)
`CLAUDE_CODE_SUBAGENT_MODEL`	Force a specific model for sub-agents	Latest (env-vars docs)
`ANTHROPIC_DEFAULT_SONNET_MODEL`	Pin the Sonnet version	Latest Sonnet (model config docs)
`ANTHROPIC_DEFAULT_OPUS_MODEL`	Pin the Opus version	Latest Opus
`ANTHROPIC_DEFAULT_HAIKU_MODEL`	Pin the Haiku version	Latest Haiku

Pin model versions on Bedrock, Vertex AI, or Foundry — alias resolution otherwise shifts when Anthropic releases new models (model config docs).

modelOverrides¶

The modelOverrides setting (v2.1.73) maps picker entries to custom provider model IDs, decoupling the UI from routing (settings docs). Combined with ANTHROPIC_DEFAULT_*_MODEL, it gives enterprise teams full control over which model serves each picker slot.

Context and Memory¶

Variable	Effect	Default
`CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`	Auto-compaction threshold (1-100)	95% (settings docs)
`CLAUDE_CODE_DISABLE_AUTO_MEMORY`	Disable auto MEMORY.md persistence	Enabled
`CLAUDE_CODE_DISABLE_1M_CONTEXT`	Remove 1M-context models from picker	Enabled (model config docs)
`CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING`	Revert to fixed `MAX_THINKING_TOKENS` budget	Adaptive
`ENABLE_CLAUDE_CODE_SM_COMPACT`	Session memory compaction (undocumented)	Disabled

Opus 4.6 and Sonnet 4.6 support 1M context via the context-1m-2025-08-07 beta header (Anthropic models page); long-context pricing applies above 200K. Set CLAUDE_CODE_DISABLE_1M_CONTEXT to suppress 1M models from the picker.

Agent and Workflow¶

Variable	Effect	Default
`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS`	Enable agent teams	Disabled (settings docs)
`CLAUDE_CODE_PLAN_MODE_REQUIRED`	Auto-set when spawning teammates; flags Plan Mode read-only	Auto-set
`CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR`	Reset to project dir after each bash command	Not set (env-vars docs)
`ENABLE_LSP_TOOL`	LSP integration: goToDefinition, findReferences, hover, documentSymbol, getDiagnostics	Disabled (issue #15619, ClaudeLog)

The LSP tool requires the marketplace plugin (/plugin marketplace add Piebald-AI/claude-code-lsps) plus a language server binary (ClaudeLog). See Semantic Context Loading for how LSP-backed tools enable semantic queries.

Tool Configuration¶

Variable	Effect	Default
`CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY`	Max concurrent tool executions	10 (env vars gist)
`BASH_MAX_OUTPUT_LENGTH`	Max bash output chars (middle-truncated)	Not specified (env-vars docs)
`CLAUDE_CODE_GLOB_TIMEOUT_SECONDS`	Glob timeout	20s (60s on WSL)

Observability¶

Variable	Effect	Default
`CLAUDE_CODE_ENABLE_TELEMETRY`	Enable OTEL collection (required before exporters)	Disabled (settings docs)
`OTEL_EXPORTER_OTLP_ENDPOINT`	OTEL endpoint for traces, metrics, logs	Not set
`OTEL_LOG_USER_PROMPTS`	Include user prompts in OTEL logs	Disabled (env vars gist)
`OTEL_LOG_TOOL_CONTENT`	Include tool content in OTEL logs	Disabled

Setting Variables¶

Set variables in settings.json under the env key (settings docs), or per-session: CLAUDE_CODE_EFFORT_LEVEL=low claude. For fleet distribution, see the managed settings drop-in; --bare mode reads the same env block in CI.

When the Prefix Taxonomy Breaks Down¶

The prefix signal is a heuristic, not a contract — Anthropic's settings docs do not define a formal naming policy. Do not rely on the prefix alone to infer stability:

Exceptions. MAX_THINKING_TOKENS, CLAUDE_AUTOCOMPACT_PCT_OVERRIDE, BASH_MAX_OUTPUT_LENGTH, and the OTEL_* family carry no tier prefix but are as load-bearing as any stable knob.
Undocumented ENABLE_*. Flags like ENABLE_CLAUDE_CODE_SM_COMPACT are absent from the official reference; name, defaults, and behavior can change without a deprecation window.
Promotion churn. Experimental flags that graduate to default are sometimes renamed or folded into settings.json fields — pinning a specific ENABLE_* in a managed drop-in can silently stop working after an upgrade. Treat the release notes as authoritative, not the prefix.

Example¶

A settings.json for a cost-conscious team deploying through Bedrock, with telemetry enabled and context compaction tuned:

{
  "env": {
    "CLAUDE_CODE_EFFORT_LEVEL": "low",
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "80",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-4-20250514",
    "CLAUDE_CODE_ENABLE_TELEMETRY": "1",
    "OTEL_EXPORTER_OTLP_ENDPOINT": "http://localhost:4317"
  },
  "modelOverrides": {
    "sonnet": "arn:aws:bedrock:us-east-1:123456789:inference-profile/sonnet-4"
  }
}

CLAUDE_CODE_EFFORT_LEVEL=low reduces reasoning depth for routine tasks. CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80 compacts earlier, keeping context from saturating. The model pin prevents silent upgrades when Anthropic ships a new Sonnet.

Key Takeaways¶

ENABLE_* flags are experimental opt-ins, DISABLE_* flags suppress defaults, CLAUDE_CODE_* flags tune stable behavior — but the convention is a heuristic, not a contract
Pin ANTHROPIC_DEFAULT_*_MODEL on Bedrock, Vertex AI, and Foundry to prevent silent model upgrades when aliases shift
CLAUDE_CODE_EFFORT_LEVEL and CLAUDE_AUTOCOMPACT_PCT_OVERRIDE are the two highest-leverage cost knobs — lowering effort and compacting earlier curbs token spend without disabling capability
CLAUDE_CODE_ENABLE_TELEMETRY=1 is the prerequisite for any OTEL export — without it the exporter env vars are ignored
Treat the release notes as authoritative for flag promotion and rename churn — pinning a specific ENABLE_* in a managed drop-in can silently stop working after an upgrade