Configuration

Configuration uses environment variables. Copy .env.example to .env; the values below are the development defaults from that file.

Application and Data

Variable	Default	Description
APP_PORT	9090	Gateway HTTP port
DB_HOST, DB_PORT	localhost, 5432	PostgreSQL connection
DB_USER, DB_PASSWORD, DB_NAME	omnidev, omnidev, omnidev	PostgreSQL credentials and database
DB_SSLMODE	disable	PostgreSQL SSL mode
REDIS_HOST, REDIS_PORT	localhost, 6379	Redis connection
REDIS_PASSWORD, REDIS_DB	empty, 0	Redis authentication and database
JWT_SECRET	development value	JWT signing secret; replace it outside local development
MINIO_ENDPOINT	localhost:9000	Attachment storage endpoint; omit MinIO settings to disable uploads
MINIO_ACCESS_KEY, MINIO_SECRET_KEY	minioadmin	MinIO credentials
MINIO_USE_SSL	false	Use TLS for MinIO
TIKA_ENDPOINT	http://localhost:9998	Tika parser for attachments and knowledge documents
TIKA_TIMEOUT	30	Parser timeout in seconds

OAuth is optional and uses the GITHUB_* and GOOGLE_* variables in .env.example.

AI Providers

Variable	Description
OPENAI_API_KEY, OPENAI_BASE_URL, OPENAI_MODELS	OpenAI or OpenAI-compatible provider settings
DEEPSEEK_API_KEY, DEEPSEEK_MODELS	DeepSeek settings
QWEN_API_KEY	Qwen provider key
ANTHROPIC_API_KEY	Anthropic provider key
AI_DEFAULT_MODEL	Default chat model (deepseek-chat in the example)
ENCRYPTION_KEY	Encrypts per-user BYOK provider keys when set

Only providers with keys are registered. Users can also manage their own provider configurations through the web UI and /user/ai-configs API.

Knowledge and RAG

Variable	Default	Description
RAG_EMBEDDING_PROVIDER	openai	openai, gemini, or ollama
RAG_EMBEDDING_MODEL	text-embedding-3-small	Embedding model
RAG_EMBEDDING_API_KEY	empty	Embedding provider key
RAG_EMBEDDING_BASE_URL	empty	Custom embedding endpoint
KNOWLEDGE_STORAGE_TYPE	local	local or minio
KNOWLEDGE_LOCAL_PATH	./data/knowledge	Local document directory
KNOWLEDGE_BUCKET	knowledge	MinIO bucket
KNOWLEDGE_PARSER_TYPE	tika	Document parser
KNOWLEDGE_CACHE_TTL	10m	Search cache TTL
KNOWLEDGE_EMBEDDING_DIMENSIONS	0	0 infers dimensions from the model response

Dimensions are tracked per knowledge base. PostgreSQL must provide the vector extension.

External MCP Tools

MCP_ENABLED=true
MCP_SERVERS='[{"name":"knowledge","url":"http://localhost:9100/mcp","api_key":"","timeout_seconds":15,"headers":{}}]'

MCP_SERVERS is a JSON array whose entries support name, url, api_key, timeout_seconds, and headers. The gateway uses streamable HTTP JSON-RPC 2.0. A server that cannot initialize is logged and skipped.

Chat runs at most five tool rounds and truncates each tool output to 8000 runes. Used tool names appear in message metadata as tools_used. MCP function calling currently applies to OpenAI, DeepSeek, and Qwen; it is not yet available for Anthropic models.

Background Agent Sandbox

Variable	Default	Description
SANDBOX_ENABLED	true	Register agent routes when Docker is available
SANDBOX_WORKSPACE_ROOT	./data/agent-workspaces	Gateway-side workspace root
SANDBOX_HOST_WORKSPACE_ROOT	empty	Host path of that same directory when the gateway runs in Docker
SANDBOX_NETWORK	none	Sandbox container network
SANDBOX_CPU_LIMIT	1	CPU limit
SANDBOX_MEMORY_LIMIT	512m	Memory limit
SANDBOX_TIMEOUT	120s	Code execution timeout
SANDBOX_PYTHON_IMAGE	python:3.12-alpine	Python runtime
SANDBOX_NODE_IMAGE	node:22-alpine	JavaScript/Node.js runtime
SANDBOX_GO_IMAGE	golang:1.22-alpine	Go runtime
SANDBOX_SHELL_IMAGE	alpine:3.19	Shell runtime
AGENT_MODEL	gpt-4o-mini	Planner model
AGENT_BASE_URL, AGENT_API_KEY	empty	Planner endpoint and key; fall back to OPENAI_*

Run make sandbox-images to pre-pull the runtime images. When the gateway is containerized, /var/run/docker.sock must be mounted and SANDBOX_HOST_WORKSPACE_ROOT must be a host path, not a path inside the gateway container.