Environment Variables¶

All variables are set in compose.yaml per service. Agents read them via config/config.py.

Agent configuration¶

Variable	Default	Description
`MODEL_NAME`	`qwen2.5:3b`	LLM model name
`OLLAMA_HOST`	`http://<project>-ollama:11434`	Ollama server URL
`LLM_PROVIDER`	`ollama`	Provider: ollama, openai, gemini, grok, anthropic, bedrock, azure
`LLM_API_KEY`	—	API key for cloud providers
`LLM_TEMPERATURE`	`0.1`	LLM temperature
`AGENT_PORT`	auto-assigned	A2A protocol port
`WEB_INTERFACE_PORT`	auto-assigned	HTTP/SSE port (if web interface enabled)
`LOG_LEVEL`	`INFO`	DEBUG, INFO, WARNING, ERROR

Semantic Layer¶

Variable	Default	Description
`SEMANTIC_LAYER_HOST`	`http://<project>-semantic-layer:10100`	Semantic Layer URL
`MCP_HOST`	`<project>-semantic-layer`	MCP server hostname
`MCP_PORT`	`10100`	MCP server port
`MCP_TRANSPORT`	`streamable-http`	Transport: streamable-http, sse
`WEAVIATE_URL`	`http://<project>-weaviate:8080`	Weaviate URL
`EMBEDDING_MODEL`	`nomic-embed-text:v1.5`	Model for generating embeddings

Security¶

Variable	Default	Description
`GUARDIAN_URL`	`http://<project>-guardian:11438`	Guardian service URL
`OPA_URL`	`http://<project>-opa:8181`	OPA policy engine URL
`A2A_VALIDATION_MODE`	`permissive`	strict, permissive, or disabled
`A2A_ENABLE_AUDIT_LOG`	`true`	Log all A2A validation attempts
`MAX_RISK_THRESHOLD`	`0.7`	Block requests above this risk score
`VALIDATION_MODE`	`permissive`	Semantic Layer validation mode
`REQUIRE_USER_VALIDATION`	`false`	Require user email for MCP calls

Artifact Store (MinIO)¶

Variable	Default	Description
`ARTIFACT_ENDPOINT`	`http://<project>-minio:9000`	MinIO URL
`ARTIFACT_ACCESS_KEY`	`minioadmin`	MinIO access key
`ARTIFACT_SECRET_KEY`	`minioadmin`	MinIO secret key
`ARTIFACT_BUCKET`	`abi-artifacts`	Default bucket
`LOG_TO_ARTIFACT_STORE`	`true`	Upload execution logs to MinIO
`LOG_AGENT_NAME`	agent name	Identifier in log bucket
`LOG_BUCKET`	`abi-logs`	Bucket for logs

Agent Memory (Redis AMS)¶

Note

Alpha. The Agent Memory Server is provisioned by the ABI Swarm scaffolding, which is under active development. These variables may change between releases.

The swarm scaffolding provisions a Redis 8 instance and a Redis Agent Memory Server (AMS) for system-wide short-term (working) and long-term memory. Agents reach it over the Docker network.

Variable	Default	Description
`AGENT_MEMORY_URL`	`http://<project>-agent-memory:8000`	AMS base URL (consumed by agents/ephemerals)
`CONTEXT_ID`	—	Session/context id used as the AMS `session_id`

AMS service configuration (set on the <project>-agent-memory service in compose.yaml):

Variable	Default	Description
`REDIS_URL`	`redis://<project>-redis-stack:6379`	Redis 8 backing store
`DISABLE_AUTH`	`true`	Disable auth (local/dev only)
`LONG_TERM_MEMORY`	`true`	Enable long-term (vector) memory
`GENERATION_MODEL`	`ollama/qwen2.5:3b`	LLM for summarization/extraction (via LiteLLM)
`FAST_MODEL`	`ollama/qwen2.5:3b`	Model for topic extraction / NER
`SLOW_MODEL`	`ollama/qwen2.5:3b`	Model for complex tasks
`EMBEDDING_MODEL`	`ollama/nomic-embed-text:v1.5`	Embedding model (768 dims)
`OLLAMA_API_BASE`	`http://ollama:11434`	Ollama endpoint for LiteLLM
`REDISVL_VECTOR_DIMENSIONS`	`768`	Vector dimensions (must match embedding model)

Note: AMS requires Redis 8+ because it uses the HSETEX command (introduced in Redis 8.0). Redis 8 also bundles RediSearch/RedisJSON in the core, so the redis:8 image is sufficient — no separate redis-stack modules needed.

Session Management¶

Note

Alpha. Framework-managed sessions are under active development. These variables may change between releases.

Sessions tie an opaque token to an internal context_id and its conversation context. The backend is pluggable: memory (per-pod, dev) or redis (shared state, LB/multi-pod safe). See Sessions & Multi-turn.

Variable	Default	Description
`SESSION_BACKEND`	`memory`	`memory` (per-pod, in-process) or `redis` (shared, multi-pod safe)
`SESSION_TTL`	`3600`	Session/context lifetime in seconds
`SESSION_REDIS_URL`	falls back to `REDIS_URL`, then `redis://localhost:6379/0`	Redis URL for the `redis` backend
`ABI_SESSION_REQUIRED`	`false`	If `true`, the orchestrator rejects `/stream` requests without a valid token

Multi-pod? Use SESSION_BACKEND=redis. With memory, each replica keeps its own sessions in RAM, so a request that lands on another pod loses the conversation. Redis shares the state across pods.

Ollama control¶

Variable	Default	Description
`START_OLLAMA`	`false`	Start local Ollama in container (distributed mode)
`LOAD_MODELS`	`false`	Pull models on container start

Agent card¶

Variable	Default	Description
`AGENT_CARD`	`./agent_cards/<name>_agent.json`	Path to agent card file