CONCEPT Two ways to set an env var — and where they matter
Environment variables configure authentication, provider routing, model choice, timeouts, and telemetry. You can set them two ways, and both feed the same variables:
- In your shell, before launching —
export ANTHROPIC_MODEL=...; claude - Inside
settings.jsonunder theenvkey — applied to every session automatically (see 2.2)
🔑 The
env block in settings.json is the durable, shareable way; a shell
export is the ad-hoc, one-terminal way. Managed policy can also pin env vars org-wide.CONCEPT The variables worth memorising
Grouped by what they control:
| Variable | Group | Purpose |
|---|---|---|
ANTHROPIC_API_KEY | Auth | API key used when authenticating to the Anthropic API |
ANTHROPIC_AUTH_TOKEN | Auth | Bearer auth token (sent as Authorization) |
ANTHROPIC_MODEL | Model | Overrides the model Claude Code uses |
CLAUDE_CODE_USE_BEDROCK | Provider | Route through Amazon Bedrock instead of the Anthropic API |
CLAUDE_CODE_USE_VERTEX | Provider | Route through Google Vertex AI |
MAX_THINKING_TOKENS | Reasoning | Cap on extended-thinking tokens (0 disables thinking) |
BASH_DEFAULT_TIMEOUT_MS | Tools | Default timeout for Bash tool commands |
DISABLE_TELEMETRY | Privacy | Turn off telemetry reporting |
DISABLE_AUTOUPDATER | Updates | Stop Claude Code from auto-updating |
CLAUDE_CODE_API_KEY_HELPER_TTL_MS | Auth | How often the apiKeyHelper value is refreshed |
NO_COLOR / FORCE_COLOR | Display | Disable / force ANSI color output |
⚠️
CLAUDE_CODE_USE_BEDROCK and CLAUDE_CODE_USE_VERTEX are the provider
switches. When set, Claude Code authenticates to that cloud (AWS / GCP) instead of the Anthropic API directly,
so credentials come from that provider, not ANTHROPIC_API_KEY.DIAGRAM Which provider does a request hit?
flowchart TD
A["claude launches"] --> B{"CLAUDE_CODE_USE_BEDROCK=1?"}
B -->|"yes"| BR["Amazon Bedrock
(AWS credentials)"] B -->|"no"| C{"CLAUDE_CODE_USE_VERTEX=1?"} C -->|"yes"| VX["Google Vertex AI
(GCP credentials)"] C -->|"no"| D{"ANTHROPIC_API_KEY set?"} D -->|"yes"| API["Anthropic API
(direct)"] D -->|"no"| LOGIN["Interactive login /
apiKeyHelper output"] classDef prov fill:#cfe2ff,stroke:#3d7bfe,color:#0a3a9a; classDef ok fill:#d4edda,stroke:#28a745,color:#155724; class BR,VX prov; class API,LOGIN ok;
(AWS credentials)"] B -->|"no"| C{"CLAUDE_CODE_USE_VERTEX=1?"} C -->|"yes"| VX["Google Vertex AI
(GCP credentials)"] C -->|"no"| D{"ANTHROPIC_API_KEY set?"} D -->|"yes"| API["Anthropic API
(direct)"] D -->|"no"| LOGIN["Interactive login /
apiKeyHelper output"] classDef prov fill:#cfe2ff,stroke:#3d7bfe,color:#0a3a9a; classDef ok fill:#d4edda,stroke:#28a745,color:#155724; class BR,VX prov; class API,LOGIN ok;
CODE Two ways to set the same variables
Shell exports for a one-off session vs. the durable env block that applies to every launch.
# --- Option A: shell, one session only --- export ANTHROPIC_MODEL=claude-opus-4-5 export MAX_THINKING_TOKENS=31999 export BASH_DEFAULT_TIMEOUT_MS=120000 claude # --- Option B: route through Amazon Bedrock instead --- export CLAUDE_CODE_USE_BEDROCK=1 export AWS_REGION=us-east-1 claude
// --- Option C: durable, every session (settings.json) --- { "env": { "ANTHROPIC_MODEL": "claude-opus-4-5", "MAX_THINKING_TOKENS": "31999", "DISABLE_TELEMETRY": "1" } }
🔑 In the
env block every value is a string — even numeric ones like
"31999" and boolean-ish flags like "1".TRADE-OFFS Shell export vs. settings.json env
✅ settings.json env
- Applies to every session automatically
- Committed & shared (project scope) or org-pinned (managed)
- Reviewable in version control
⚠️ Shell export
- Fast, no file edit — great for one-off experiments
- But: vanishes with the terminal; easy to forget
- Not shared, so teammates don't get it
EXAM TRAPS
Trap: "
ANTHROPIC_MODEL is a settings.json key." → It's an environment
variable. The settings.json key for the model is just model (2.2).Trap: "Set
CLAUDE_CODE_USE_BEDROCK and supply
ANTHROPIC_API_KEY to use Bedrock." → On Bedrock, auth comes from AWS credentials, not the
Anthropic API key. The two provider switches redirect where requests go.Trap: "
MAX_THINKING_TOKENS=0 gives unlimited thinking." → 0
disables extended thinking. A higher number raises the cap.Trap: "Env values in settings.json can be raw numbers." → They must be
quoted strings in JSON.
PRACTICE
Q1 · You want every session in this repo to route through Vertex AI. What do you set, and where's the durable place?
Set CLAUDE_CODE_USE_VERTEX (to "1"). Durably, put it in the
env block of .claude/settings.json so it applies on every launch; ad-hoc, you'd
export it in your shell.
Q2 · Which variable, set to 0, turns off extended thinking?
MAX_THINKING_TOKENS. Setting it to 0 disables thinking; any positive
value caps the thinking-token budget.
Q3 · Name the variable that stops Claude Code from phoning home with telemetry.
DISABLE_TELEMETRY (set to 1). You can put it in the shell or in the
settings.json env block.