Configuration

⚙️ Configuration

The tables below separate application defaults from the effective values in the shipped Compose profile.

Variable	Compose value	Note
ADMIN_TOKEN	change-me	Unsafe placeholder; replace it before first use as shown above
PROXY_API_KEY	empty unless set in .env	Empty leaves inference routes open
REQUEST_BODY_LIMIT	500mb	Overrides the 100mb application default
MODELS_CLIENT_VERSION	1.0.0	Overrides the application default
PROXY_MODELS	gpt-5.3-codex,gpt-5.2-codex,gpt-5-codex	Fallback model catalog
EXCLUDED_PROVIDER_MODELS	openai-compatible:gpt-5.5,mistral:mistral-medium-latest	Overrides the empty application default

Compose uses .env for interpolation only. Variables not listed under the service's environment section must be added there or supplied by an override file before they reach the container.

Core settings — ports, storage, credentials, and models

Core settings

Variable	Default	Purpose
MULTIVIBE_CONTROL_PLANE	false (true in the container)	Make Node serve only the control plane while Rust owns /v1
PORT	1455	Node port in single-process mode; the Compose profile overrides it to 1456
HOST	Node.js all-interface default	Node listen address; the native control plane pins literal loopback
CONTROL_PLANE_PORT	1456	Loopback Node control-plane port
V1_EDGE_HOST	0.0.0.0	Native Rust edge listen address
V1_EDGE_PORT	1455	Native Rust edge listen port
V1_EDGE_BASE_URL	http://127.0.0.1:1455	Internal URL used by control-plane jobs and host integrations
NODE_CONTROL_PLANE_URL	http://127.0.0.1:1456	Loopback target for non-/v1 edge fallback requests
V1_EDGE_INTERNAL_JOB_TOKEN	generated by the container entrypoint	Shared capability for control-plane job execution; never expose it publicly
ADMIN_TOKEN	empty	Dashboard/admin secret; empty disables the admin check
PROXY_API_KEY	empty	Shared inference/API key
PROXY_API_KEYS	empty	JSON object mapping application names to keys
STORE_PATH	/data/accounts.json	Account and dashboard-managed state
PROVIDER_AGENT_ENABLED	false	Supervise the packaged provider agent when explicitly enabled
PROVIDER_AGENT_BINARY	/opt/multivibe/bin/multivibe-provider-agent	Absolute packaged agent binary path
PROVIDER_AGENT_STATE_PATH	beside STORE_PATH	Local explicit provider model selection
PROVIDER_AGENT_RUNTIME_STATE_PATH	beside STORE_PATH	Manual loopback endpoints and optional runtime bearers
PROVIDER_AGENT_DEVICE_KEY_PATH	beside STORE_PATH	Ed25519 device identity and relay-shadow sequence
PROVIDER_AGENT_ENROLLMENT_STATE_PATH	beside STORE_PATH	Submitted Cloud shadow node view without the grant
PROVIDER_AGENT_CLOUD_API_URL	https://auth.multivibe.cloud	Fixed Cloud enrollment API origin
OAUTH_STATE_PATH	/data/oauth-state.json	OAuth state
TRACE_FILE_PATH	/data/requests-trace.jsonl	Recent traces
TRACE_STATS_HISTORY_PATH	/data/requests-stats-history.jsonl	Long-term lightweight usage
ANONYMOUS_USAGE_STATE_PATH	/data/anonymous-usage-state.json	Daily anonymous-sharing retry envelope
ANONYMOUS_USAGE_API_BASE_URL	https://api.multivibe.cloud	Public allowlist and anonymous daily aggregate API
CODEX_PROJECTS_PATH	/data/codex-projects.json	Codex project registry
JOBS_DB_PATH	/data/jobs.sqlite	Legacy/control-plane job database
V1_EDGE_JOBS_PATH	beside STORE_PATH (/data/v1-edge-jobs.json in Compose)	Native Rust edge deferred-job state
JOB_WORKER_CONCURRENCY	16	Global worker concurrency; destination capacity still applies
REQUEST_BODY_LIMIT	100mb	JSON or decompressed-zstd body limit
TRACE_RETENTION_MAX	1000	Recent full traces retained; minimum 100
TRACE_INCLUDE_BODY	false	Persist full request bodies in recent traces
TRACE_INCLUDE_HEADERS	false	Persist sanitized inbound headers in recent traces
CODEX_PROJECT_REGISTRATION_TOKEN	ADMIN_TOKEN	Limited project-registration secret
CODEX_SESSION_AFFINITY	false	In-memory per-session account stickiness
CODEX_SESSION_AFFINITY_TTL_MS	3600000	Affinity TTL in the native Rust edge
CODEX_SESSION_AFFINITY_MAX_ENTRIES	10000	Affinity cache limit
INFERENCE_IDEMPOTENCY_TTL_MS	300000	Completed synchronous inference replay TTL
INFERENCE_IDEMPOTENCY_IN_FLIGHT_TIMEOUT_MS	300000	Maximum single-flight reservation lifetime
INFERENCE_IDEMPOTENCY_MAX_ENTRIES	1000	Combined in-flight and completed entry limit
INFERENCE_IDEMPOTENCY_MAX_BYTES	33554432	Global completed-response memory budget
INFERENCE_IDEMPOTENCY_MAX_RESPONSE_BYTES	1048576	Per-response replay limit
PROXY_MODELS	gpt-5.3-codex,gpt-5.2-codex,gpt-5-codex	Fallback model catalog
MODELS_CLIENT_VERSION	0.144.1	Codex identity used for discovery/runtime requests
MODELS_CACHE_MS	600000	Model-catalog refresh interval
EXCLUDED_PROVIDER_MODELS	empty	Comma-separated provider:model exclusions
CLAUDE_CODE_MODEL	gpt-5.6-luna	Claude Code opus/sonnet upstream
CLAUDE_CODE_FAST_MODEL	gpt-5.4-mini	Claude Code haiku/fast upstream

Routing, cache, retry, and block tuning

Variable	Default	Purpose
USAGE_CACHE_TTL_MS	300000	Usage snapshot freshness
USAGE_REFRESH_INTERVAL_MS	600000	Background quota refresh interval when no requests are running
USAGE_TIMEOUT_MS	10000	Provider usage-probe timeout
USAGE_STALE_WHILE_REVALIDATE	true	Route with bounded stale/missing usage while refreshing
USAGE_STALE_MAX_AGE_MS	1800000	Maximum stale usage age
MODELS_STALE_WHILE_REVALIDATE	true	Serve a bounded stale model catalog while refreshing
MODELS_STALE_MAX_AGE_MS	1800000	Maximum stale catalog age
FIVE_HOUR_QUOTA_THRESHOLD_PERCENT	90	Near-limit five-hour threshold
BLOCK_FALLBACK_MS	1800000	Quota block fallback without a usable reset
RATE_LIMIT_BLOCK_MS	60000	Ordinary rate-limit account/model block
MODEL_NOT_FOUND_BLOCK_DURATION_MS	3600000	Model-not-found account/model block
MAX_ACCOUNT_RETRY_ATTEMPTS	10	Candidate accounts tried after quota failures
MAX_UPSTREAM_RETRIES	5	Same-account transient retries
UPSTREAM_BASE_DELAY_MS	2000	Retry backoff base
HANG_RETRY_INTERVAL_MS	10000	Delay while all accounts are temporarily exhausted
HANG_RETRY_MAX_DURATION_MS	120000	Maximum all-account wait
EMPTY_RESPONSE_BLOCK_THRESHOLD	3	Empty outputs before a temporary block
EMPTY_RESPONSE_BLOCK_DURATION_MS	30000	Empty-output block duration
EMPTY_RESPONSE_WINDOW_MS	300000	Empty-output counting window
TOKEN_REFRESH_MARGIN_MS	60000	OAuth refresh margin
ACCOUNT_FLUSH_INTERVAL_MS	5000	Account-store write debounce

Provider, OAuth, Realtime, and observability settings

Variable	Default
CHATGPT_BASE_URL	https://chatgpt.com
UPSTREAM_PATH	/backend-api/codex/responses
UPSTREAM_COMPACT_PATH	/backend-api/codex/responses/compact
MISTRAL_BASE_URL	https://api.mistral.ai
OPENCODE_BASE_URL	https://opencode.ai/zen
OPENCODE_CONSOLE_URL	https://opencode.ai/console
OPENCODE_OAUTH_CLIENT_ID	opencode-cli
ZAI_BASE_URL	https://api.z.ai
ZAI_MODELS_PATH	/api/paas/v4/models
XAI_BASE_URL	https://cli-chat-proxy.grok.com/v1
XAI_AUTH_PATH	$HOME/.grok/auth.json
OAUTH_REDIRECT_URI	http://localhost:1455/auth/callback
REALTIME_PROVIDER	openai
REALTIME_WEBRTC_CALL_URL	empty
REALTIME_REQUEST_TIMEOUT_MS	30000
SENTRY_DSN	empty
SENTRY_ENVIRONMENT	NODE_ENV or production
SENTRY_TRACES_SAMPLE_RATE	0.1

Provider endpoint paths, OAuth endpoints, and Grok client-identity fields can also be overridden. The authoritative definitions live in src/config.ts and src/oauth-config.ts.