Environment variables

claude-multi runtime variables

These control claude-multi itself. Set them in your shell before launching.

Variable	Default	Description
`CLAUDE_MULTI_INK`	`true`	Set to `false` to use the simpler prompts-based UI instead of the Ink TUI
`CLAUDE_MULTI_HOME`	home directory	Base directory for central state. Health status and migration backups resolve under `$CLAUDE_MULTI_HOME/.claude-multi/`
`CLAUDE_MULTI_CLAUDE_PATH`	unset	Absolute path to the `claude` binary. Overrides the PATH lookup when wrappers are created or repaired. The file must exist
`CLAUDE_MULTI_UPDATE_CHECK`	`false`	Set to `true` to check for claude-multi and Claude Code updates on launch

Examples:

# Use the simpler UI
CLAUDE_MULTI_INK=false claude-multi

# Point wrapper creation at a specific Claude Code install
CLAUDE_MULTI_CLAUDE_PATH=/opt/claude/bin/claude claude-multi add glm

# Store central state under a test directory
CLAUDE_MULTI_HOME=/tmp/test-env claude-multi doctor check

CLAUDE_MULTI_HOME does not move an instance’s config directory. Each instance still lives at ~/.claude-<name>/ unless you pass --config to add. The instance registry still resolves from your home directory, so this variable is mainly useful for tests and isolated health or migration runs.

Claude Code instance variables

These live in each instance’s settings.json under the env key. Provider templates fill them in. You can edit them by hand or let claude-multi doctor fix refresh them from the current template.

Core API variables

Variable	Description	Example
`ANTHROPIC_AUTH_TOKEN`	Bearer token for the `Authorization` header. Provider API keys go here	`sk-your-key-here`
`ANTHROPIC_BASE_URL`	Provider’s Anthropic-compatible API endpoint	`https://api.deepseek.com/anthropic`

Model mapping variables

Variable	Description	Maps to
`ANTHROPIC_MODEL`	Primary model for the session	The main conversation
`ANTHROPIC_SMALL_FAST_MODEL`	Deprecated. Use `ANTHROPIC_DEFAULT_HAIKU_MODEL` for the fast model	Superseded by `ANTHROPIC_DEFAULT_HAIKU_MODEL`; current templates do not write it
`ANTHROPIC_DEFAULT_SONNET_MODEL`	Sonnet-tier model	Used when Claude Code requests a “sonnet” class model
`ANTHROPIC_DEFAULT_OPUS_MODEL`	Opus-tier model	Used when Claude Code requests an “opus” class model
`ANTHROPIC_DEFAULT_FABLE_MODEL`	Fable-tier model	Used by the `fable` alias and automatic fallback. Without it, Claude Code asks the provider for `claude-fable-5-1`, which a third-party endpoint cannot serve
`ANTHROPIC_DEFAULT_HAIKU_MODEL`	Haiku-tier model	Used when Claude Code requests a “haiku” class model, and by most templates for the fast slot

Sub-agent, effort, and thinking variables

Variable	Description	Example
`CLAUDE_CODE_SUBAGENT_MODEL`	Default model for subagents, agent teams, and workflow agents. A per-invocation model or an agent definition’s `model` field can override it	`deepseek-flash`
`CLAUDE_CODE_EFFORT_LEVEL`	Reasoning effort for the main model. Accepted values depend on the model	`low`, `medium`, `high`, `xhigh`, `max`
`CLAUDE_CODE_ALWAYS_ENABLE_EFFORT`	Set to `1` to send the effort parameter for custom model IDs. Templates that set an effort level include it because Claude Code cannot infer effort support from a third-party model name	`1`
`MAX_THINKING_TOKENS`	Fixed thinking budget in tokens. Applies only to models that accept a budget; adaptive-reasoning models and providers that use an effort parameter can ignore it	`32000`

Output variables

Claude Code caps output at 128,000 tokens for an unrecognized model ID and defaults to 32,000 when the variable is unset. A model that lists a larger maximum output cannot exceed the Claude Code cap.

Variable	Description	Example
`CLAUDE_CODE_MAX_OUTPUT_TOKENS`	Maximum response tokens	`128000`

Timeout and traffic variables

Variable	Description	Example
`API_TIMEOUT_MS`	Time in milliseconds each request attempt has to receive response headers. It does not cap how long a streaming response runs	`3000000`
`CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC`	Disables nonessential traffic. The GLM and MiniMax templates set it to `1`	`1`

Context window and auto-compaction

Claude Code assumes a 200K context window for an unrecognized model ID. When the real window differs, tell Claude Code with the [1m] suffix on the model name or with CLAUDE_CODE_MAX_CONTEXT_TOKENS. Without that, compaction runs against the wrong window and the API call can fail mid-session.

Claude Code compresses older conversation history once usage crosses a share of the window. CLAUDE_CODE_AUTO_COMPACT_WINDOW sets the window used for that decision; it accepts values from 100,000 to 1,000,000. CLAUDE_AUTOCOMPACT_PCT_OVERRIDE sets the percentage that triggers compaction.

Variable	Description	Example
`CLAUDE_CODE_MAX_CONTEXT_TOKENS`	Context window override. Applies directly to unrecognized non-`claude-` IDs without `[1m]`; with `[1m]`, pair it with `CLAUDE_CODE_DISABLE_1M_CONTEXT=1`; for recognized IDs it needs `DISABLE_COMPACT`	`983616` for Qwen
`CLAUDE_CODE_AUTO_COMPACT_WINDOW`	Window used to decide when to compact, in tokens. Range is 100000 to 1000000	`786432` for DeepSeek
`CLAUDE_AUTOCOMPACT_PCT_OVERRIDE`	Percentage of the window that triggers compaction. It can only make compaction happen earlier; values above the default are ignored	`75`

The templates set this window at or below the model’s context limit. Values below 100,000 or above 1,000,000 are outside the supported range.

Per-provider defaults

Each template writes its own base URL, model slots, timeout, and tuning values. This table matches src/templates.ts in 0.13.2.

Template	Provider page	Base URL	Primary slot	Fable slot	Fast slot	Timeout
`glm`	GLM	`https://api.z.ai/api/anthropic`	`glm-5.3[1m]`	`glm-5.3[1m]`	`glm-5.3-flash[1m]`	`3000000`
`minimax`	MiniMax	`https://api.minimax.io/anthropic`	`MiniMax-M3.1-Flash[1m]`	`MiniMax-M3.1-Flash[1m]`	`MiniMax-M3.1-Flash[1m]`	`3000000`
`deepseek`	DeepSeek	`https://api.deepseek.com/anthropic`	`deepseek-flash[1m]`	`deepseek-flash[1m]`	`deepseek-flash`	`3000000`
`mimo`	Xiaomi MiMo	`https://api.xiaomimimo.com/anthropic`	`mimo-v2.6-pro[1m]`	`mimo-v2.6-pro[1m]`	`mimo-v2.6-flash[1m]`	`3000000`
`mimo-token`	Xiaomi MiMo	`https://token-plan-cn.xiaomimimo.com/anthropic`	`mimo-v2.6-pro[1m]`	`mimo-v2.6-pro[1m]`	`mimo-v2.6-flash[1m]`	`3000000`
`kimi`	Moonshot Kimi	`https://api.moonshot.ai/anthropic`	`kimi-k3.1-pro[1m]`	`kimi-k3.1-pro[1m]`	`kimi-k3.1-pro[1m]`	`3000000`
`qwen`	Alibaba Qwen	`https://dashscope-intl.aliyuncs.com/apps/anthropic`	`qwen3.8-max`	`qwen3.8-max`	`qwen3.8-flash`	`3000000`
`qwen-coding`	Alibaba Qwen	`https://coding-intl.dashscope.aliyuncs.com/apps/anthropic`	`qwen3.7-plus`	`qwen3.7-plus`	`qwen3.7-plus`	`3000000`

The primary column shows ANTHROPIC_MODEL. The Fable column shows ANTHROPIC_DEFAULT_FABLE_MODEL. The fast column shows ANTHROPIC_DEFAULT_HAIKU_MODEL. Templates that set CLAUDE_CODE_SUBAGENT_MODEL point it at the fast model.

Every template maps the fable alias to its primary model. Claude Code otherwise resolves fable to claude-fable-5-1, and a third-party Anthropic-compatible endpoint rejects that ID.

Other defaults that differ by template:

Template	Output and reasoning defaults
`glm`	`CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000`, `CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000`, `CLAUDE_CODE_EFFORT_LEVEL=high`, `CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1`, `MAX_THINKING_TOKENS=8000`, `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1`
`minimax`	`CLAUDE_CODE_AUTO_COMPACT_WINDOW=524288`, `CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000`, `CLAUDE_CODE_EFFORT_LEVEL=max`, `CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1`, `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1`
`deepseek`	`CLAUDE_CODE_AUTO_COMPACT_WINDOW=786432`, `CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000`, `CLAUDE_CODE_EFFORT_LEVEL=max`, `CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1`, `MAX_THINKING_TOKENS=32000`
`mimo`	`CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000`
`mimo-token`	`CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000`
`kimi`	`CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000`, `CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000`, `CLAUDE_CODE_EFFORT_LEVEL=max`, `CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1`, `MAX_THINKING_TOKENS=16000`
`qwen`	`CLAUDE_CODE_MAX_CONTEXT_TOKENS=983616`, `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=75`, `CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000`, `MAX_THINKING_TOKENS=16000`
`qwen-coding`	`CLAUDE_CODE_MAX_CONTEXT_TOKENS=983616`, `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=75`, `CLAUDE_CODE_MAX_OUTPUT_TOKENS=128000`, `MAX_THINKING_TOKENS=16000`

The mimo-token base URL depends on the region you picked at creation. cn uses token-plan-cn.xiaomimimo.com, sgp uses token-plan-sgp.xiaomimimo.com, and ams uses token-plan-ams.xiaomimimo.com. All three keep the /anthropic path.

Template status

The 0.13.2 templates match the current vendor model lists as of 2026-11-03. The notes below cover plan coverage and model availability where they affect a slot.

Alibaba Qwen

The qwen template maps qwen3.8-max to the primary, opus, and fable slots, qwen3.7-plus to sonnet, and qwen3.8-flash to haiku and subagents. It sets CLAUDE_CODE_MAX_CONTEXT_TOKENS=983616 and compacts at 75 percent. It does not set an effort level, so the provider default applies.

The qwen-coding template uses qwen3.7-plus in every slot. The Coding Plan includes Qwen3.7-Plus but not Qwen3.8-Max or Qwen3.8-Flash. That is a plan limit, not an endpoint limit. If the plan rejects requests above 200K, lower CLAUDE_CODE_MAX_CONTEXT_TOKENS to 200000.

qwen3-coder-next and qwen3-coder-plus are absent from the current Anthropic-compatible model matrix. Only qwen3-coder-flash remains as a legacy Coder model.

MiniMax

The minimax template uses MiniMax-M3.1-Flash[1m] in every slot, including Fable and subagents, with a 524288 auto-compact window. It sets effort to max and adds CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1. M3 is the previous generation and remains available. MiniMax says a Pro model will follow, but no ID or date is published.

DeepSeek

The deepseek template uses deepseek-flash[1m] for primary, opus, sonnet, and Fable, and deepseek-flash for haiku and subagents, with a 786432 auto-compact window. It sets effort to max and adds CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1. deepseek-flash serves V4.1-Flash-1017. deepseek-v4-pro is V4-Pro-0813, the previous generation, and remains available as an optional model. There is no released V4.1 Pro.

DeepSeek’s Anthropic endpoint ignores budget_tokens, so MAX_THINKING_TOKENS=32000 is a client-side hint, not a provider budget. Thinking depth follows CLAUDE_CODE_EFFORT_LEVEL.

Xiaomi MiMo

The mimo and mimo-token templates use mimo-v2.6-pro[1m] for primary, opus, sonnet, and Fable, and mimo-v2.6-flash[1m] for haiku and subagents. Both models have a 1M context window and 128K output. The templates do not set an effort level. See Xiaomi MiMo for setup details.

Moonshot Kimi

The kimi template uses kimi-k3.1-pro[1m] in every slot, including Fable. Standard speed is the default. Highspeed is optional and costs more per token. It sets effort to max and adds CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1. Kimi always thinks and has no thinking-budget parameter, so MAX_THINKING_TOKENS=16000 is only a client hint. See Moonshot Kimi for setup details.

GLM

The glm template maps glm-5.3[1m] to the primary, opus, sonnet, and Fable slots and glm-5.3-flash[1m] to haiku. It sets effort to high and adds CLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1. Z.ai’s current default configuration maps opus, sonnet, and haiku to glm-5.3-flash[1m]; the template keeps the manual split so the main slots use the flagship model. GLM-5-Turbo is not a Coding Plan model and is not a template tier.

If you created an instance before 0.13.2 (including 0.13.0 and 0.13.1), run claude-multi doctor check, then claude-multi doctor fix. The sync updates model slots, the Fable alias, the effort opt-in, and known legacy defaults while keeping your API key and custom values.

Where to set variables

claude-multi variables go in your shell profile (.zshrc, .bashrc, .profile) or in the environment before you run claude-multi.

Instance variables go in ~/.claude-<name>/settings.json under env. The provider template fills them when you create the instance, and they apply to that instance only. They do not affect other instances or your default ~/.claude setup. See Configuration for the file layout and Plugins and MCP for MCP server settings.

Reviewed 2026-11-03