Environment Variables Reference

Complete reference of runtime environment variables used by the current codebase.

Core Application

Variable	Default	Description
ENVIRONMENT	development	Runtime environment (development, staging, production)
DEBUG	false	Enable debug mode (verbose logging, auto-reload)
LOG_LEVEL	INFO	Logging level: DEBUG, INFO, WARNING, ERROR
SYSTEM_MODE	real	Runtime guardrail mode (dummy, dry_run, real)

LLM Provider

Variable	Default	Description
LLM_BACKEND	vllm	Inference backend (vllm or ollama)
VLLM_MODEL	granite-soc-specialist	vLLM specialist model
VLLM_ORCHESTRATOR_MODEL	granite-soc-specialist	vLLM orchestrator model
VLLM_BASE_URL	http://vllm:8000/v1	vLLM instance URL
OLLAMA_MODEL	granite4:8b	Ollama specialist model
OLLAMA_ORCHESTRATOR_MODEL	granite4:8b	Ollama orchestrator model. Keep this equal to OLLAMA_MODEL for single-model local deployments.
OLLAMA_BASE_URL	http://ollama:11434	Ollama instance URL

Local Single-Model Controls

These variables keep all AuroraSOC agents on one local inference service and one model tag, which is the recommended MVP path for 8 GB VRAM laptops and other constrained single-machine deployments.

Variable	Default	Description
GRANITE_SINGLE_MODEL_MODE	true	Forces Ollama specialist and orchestrator agents to resolve to the same model tag, even if per-agent model routing or a separate orchestrator tag is configured.
GRANITE_USE_SHARED_MODEL_POOL	true	Reuses process-local BeeAI ChatModel instances for identical backend/model/provider tuples.
GRANITE_MAX_CONCURRENT_REQUESTS	1	Recommended local inference concurrency. Keep at 1 on 8 GB VRAM to avoid model thrash.
GRANITE_REQUEST_QUEUE_SIZE	16	Planned bounded queue size for shared LLM request admission.
GRANITE_INFERENCE_TIMEOUT_SECONDS	180	Planned per-request timeout budget for local LLM inference.

For a quota-limited local install, inspect installed models first with ollama list. If a usable model is already present, point both OLLAMA_MODEL and OLLAMA_ORCHESTRATOR_MODEL at that tag instead of pulling another model.

PostgreSQL

Variable	Default	Description
PG_HOST	postgres	PostgreSQL hostname
PG_PORT	5432	PostgreSQL port
PG_USER	aurorasoc	Database user
PG_PASSWORD	(required in compose)	Database password
PG_DATABASE	aurorasoc	Database name

Redis

Variable	Default	Description
REDIS_URL	redis://redis:6379	Direct connection URL

NATS JetStream

Variable	Default	Description
NATS_URL	nats://localhost:4222	NATS server URL
NATS_STREAM_NAME	AURORA	JetStream stream name

MQTT

Variable	Default	Description
MQTT_HOST	mosquitto	MQTT broker hostname
MQTT_PORT	8883	MQTT broker port (mTLS-first)
MQTT_USERNAME	—	MQTT username (optional)
MQTT_PASSWORD	—	MQTT password (optional)
MQTT_TOPIC_PREFIX	aurora	MQTT topic prefix

pgvector (PostgreSQL Extension)

Vector embedding storage is handled by pgvector inside the same PostgreSQL instance. No separate service is needed. Relevant pool and SSL settings live in the PostgreSQL section above.

Variable	Default	Description
PG_POOL_SIZE	20	SQLAlchemy connection pool size
PG_MAX_OVERFLOW	10	Extra connections beyond pool size
PG_SSLMODE	prefer	SSL mode (disable, prefer, require, verify-ca, verify-full)

HashiCorp Vault

Variable	Default	Description
VAULT_ADDR	http://vault:8200	Vault server URL
VAULT_TOKEN	—	Vault access token
VAULT_KV_MOUNT	secret	KV-v2 secrets mount path
VAULT_PKI_MOUNT	pki_iot	PKI mount path

Authentication

Variable	Default	Description
JWT_SECRET_KEY	—	Required. JWT signing secret (≥32 chars)
JWT_EXPIRY_HOURS	24	Token lifetime in hours
API_SERVICE_KEY	—	Required. static bootstrap API key for service auth

Approval Policy

Variable	Default	Description
APPROVAL_EXPIRES_MINUTES	30	Default expiration window for human approval requests
APPROVAL_WAIT_TIMEOUT_SECONDS	300	Default wait timeout for approval polling flows
APPROVAL_POLL_INTERVAL_SECONDS	5	Poll interval for approval decision checks

Observability

Variable	Default	Description
OTEL_EXPORTER_ENDPOINT	http://otel-collector:4317	OTLP gRPC endpoint
OTEL_SERVICE_NAME	aurorasoc	Service name in traces
OTEL_PROMETHEUS_PORT	9090	Prometheus metrics port used by config

A2A Resolution

Used to override agent service discovery statically:

Variable	Default	Description
A2A_DISCOVERY_MODE	compose	How to resolve agent IP (compose or k8s)
A2A_CLIENT_HOST	—	Global override for all A2A target hostnames

Agent Deployment

Variable	Default	Description
ENABLED_AGENTS	all	Comma-separated list of specialist agent names to instantiate at startup, or all to start the full fleet (e.g. SecurityAnalyst,ThreatHunter,IncidentResponder). The Orchestrator always runs regardless of this setting.

GPU / vLLM Tuning

These variables are used by the vllm service in docker-compose.gpu.yml when running on consumer GPUs with limited VRAM.

Variable	Default	Description
VLLM_GPU_MEMORY_UTIL	0.90	Fraction of GPU VRAM reserved for vLLM's KV-cache (0.0–1.0). Reduce if OOM.
VLLM_MAX_MODEL_LEN	8192	Maximum sequence length (prompt + completion tokens). Lower = less VRAM.
CUDA_VISIBLE_DEVICES	(unset)	GPU index to expose to the vLLM container. Leave unset to use all GPUs.

Quick Validation

Check all loaded settings by starting the API with DEBUG=true:

DEBUG=true python -m aurorasoc.api.main

Settings will be printed at INFO level during startup.