Claude Code environment variables: the useful ones

Claude Code environment variables decide which credential it signs in with (ANTHROPIC_API_KEY), which model it runs (ANTHROPIC_MODEL), where requests go (ANTHROPIC_BASE_URL, HTTPS_PROXY) and what it reports (DISABLE_TELEMETRY). Export one in your shell for that terminal, or put it in the env block of ~/.claude/settings.json for every run.

Anthropic's reference lists well over a hundred. The tables below hold the ones people actually set, grouped by what they change, as the reference read on 22 October 2026. Variables are added, renamed and retired between releases, so if one here does nothing, run claude --version and look it up in the reference again.

Where to set them#

You can use the shell and a settings file together. A shell export lasts for the terminal it was typed in:

The env block of a settings file reaches every claude run that reads the file:

{
  "env": {
    "API_TIMEOUT_MS": "1200000",
    "BASH_DEFAULT_TIMEOUT_MS": "300000"
  }
}

Each settings file reaches different people.

Which value wins#

• A variable set in both your shell and a settings env block: the settings value usually applies, because Claude Code overwrites the inherited one.
• Between files, managed settings beat user and project settings.
• A settings file can set a variable but not remove one. To cancel one your shell exports, set it to an empty string: "CLAUDE_CODE_USE_VERTEX": "".
• Project and local settings can't set a few variables, CLAUDE_CONFIG_DIR among them.

Values in JSON are always strings: write "1", not 1.

Model and reasoning#

The three ANTHROPIC_DEFAULT_*_MODEL variables matter most on Bedrock and Vertex, where Anthropic says to pin every model before a team rollout. Our Bedrock and Vertex guide has the IDs.

CLAUDE_CODE_MAX_OUTPUT_TOKENS is the one to reach for when a reply stops with an output token error. Our guide to Claude's response exceeding the output token maximum says how high it can go on each model.

Sign-in and accounts#

One forgotten export ANTHROPIC_API_KEY=... line in ~/.zshrc is enough for Claude Code to use the key: your Pro or Max plan sits unused and every turn is billed per token. /status names the credential in use. The other reasons Claude Code asks for /login again, or bills a key when you expected the plan, are in Claude Code keeps asking to log in.

Anthropic's reference calls CLAUDE_CONFIG_DIR "useful for running multiple accounts side by side": CLAUDE_CONFIG_DIR=~/.claude-work claude keeps that account's settings and history in its own folder. The two-folder setup, with shell aliases, is in switch Claude accounts without logging out.

Provider, gateway and network#

Bedrock and Vertex bring their own set (AWS_REGION, CLOUD_ML_REGION, ANTHROPIC_VERTEX_PROJECT_ID and more), covered in the Bedrock and Vertex guide. For a gateway, Anthropic's gateway guide shows how to test ANTHROPIC_BASE_URL with one curl before starting Claude Code. If requests don't get through, our guide to Claude Code's Unable to connect to API error covers a leftover ANTHROPIC_BASE_URL, HTTPS_PROXY and NODE_EXTRA_CA_CERTS.

Privacy and telemetry#

DISABLE_TELEMETRY=0 still turns telemetry off, because Claude Code only checks that the variable has a value, and DISABLE_ERROR_REPORTING=false works the same way. To opt back in, unset DISABLE_TELEMETRY. CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 also stops auto-updates, so with it set, updating is up to you.

Updates#

To confirm, run claude doctor: its Auto-updates line says disabled (set by env: DISABLE_AUTOUPDATER).

Behaviour and limits#

Check for DISABLE_PROMPT_CACHING=1 left over from debugging if your limits drain fast: with it on, every request re-reads the whole context at full price. A large MAX_MCP_OUTPUT_TOKENS lets one tool reply fill the window. How caching and context add up per turn is in why Claude Code uses so many tokens.

Check what's in effect#

• /status shows the model, the account or key in use, the base URL and, on a cloud provider, the region.
• claude doctor shows install health, settings-file errors and the update setting.
• A status line can print the model and context on every reply, so a wrong variable shows up straight away.

Accounts without the variables#

CLAUDE_CONFIG_DIR and ANTHROPIC_API_KEY are also how people keep a work plan, a personal plan and a key for when both run out apart: one terminal per folder or key. When the work account hits its limit, the task stops in that terminal, and you start it again in the other one.

Shuriken does that juggling for you. You add each account or key once as a source, and hot swapping takes it from there: Shuriken moves your chats to another of your accounts or sources when one nears its limit, inside the running chat and on every device. "Switch at" is 99% by default. An API key for the same provider stays out of the way until you choose: "Switch to API keys" is off by default, and once on, each fallback run stops at a $25 spend cap unless you change it. Only add accounts you own and are entitled to use; providers' terms may not allow sharing a subscription sign-in.

If you'd rather stop exporting variables per account, download Shuriken.