Skip to content

Installation and operations

NomOS runs on Kubernetes and is installed with a Helm chart. The reference installation runs on AWS EKS in the Zurich region. Other Kubernetes environments are intended but not yet proven as an installation. A SaaS offering operated by Ainomos is planned.

Part Job
brain-core Backend with rooms, knowledge, rules, evidence, and the MCP server
web-ui Interface for chat, rooms, rules, inbox, and administration
Worker Scheduled jobs such as Dreaming, log shipping, and maintenance
Keycloak Sign-in, or the connection to your own OIDC or SAML identity provider
OPA Access decisions for rooms
Presidio Personal data detection
LiteLLM Gateway for every model call, with cost per room
Valkey Cache

You provide a PostgreSQL 16 database with the pgvector extension.

Every model call goes through the gateway. The default is Amazon Bedrock with EU inference profiles.

  • Your own models for answers: platform admins register additional models with an OpenAI-compatible interface in the model catalogue, for example locally hosted ones. Each model carries a location (in the cluster, in your cloud, or external) and the data classes it may process.
  • Background jobs and search: Dreaming and some checks use fixed Claude models today, and search uses the Titan embedding model. Letting an installation choose its own models for these is planned.

NomOS uses the installation’s identity provider. If sign-in through the identity provider fails, a separate emergency access keeps admins from being locked out.

Signed evidence needs an Ed25519 key. Create it as a secret and manage it separately from the database. The public half is used to verify the records.