Skip to main content

Security & data boundaries

Libra OS can run inside your infrastructure. Its configuration determines which data is processed locally and which data goes to external services.

Storage and processing are separate

ComponentData it may receiveConfiguration to check
Model gatewayPrompts, retrieved passages, tool results, and generated outputAnswer, planner, and skill model settings
Embedding serviceDocument chunks and retrieval queriesEmbedding endpoint and model
Memory worker modelConversation content to summarizeMemory worker model and endpoint
Search / page-fetch providerSearch queries and requested URLsSearch and fetch backends
Tool callback / connectorInputs required by the operationCallback URL, credentials, tool policy
Runtime storesDocuments, conversations, memory, jobs, and audit recordsDatabase, filesystem, access, retention, and backups

For fully offline operation, configure local models and embeddings, use compatible local tools, and enforce the intended network boundary. An offline license does not make a hosted model available offline. See Model settings and Portability.

Identity and access

Authenticate requests with a deployment-issued credential. Agent ownership, collection restrictions, and approval-group membership serve different purposes. owner_employee supplies an agent's shared defaults; it is not a replacement for user authorization or a tenant identifier.

End-user identity overrides are accepted only for authorized integration callers. Follow Managing memory before routing several users through a shared credential.

Screening, grounding, and approvals

The firewall and configured guardrails screen requests and responses. They reduce risk; they do not prove that every output is correct or contains no sensitive information.

Inspect grounding and source evidence when the application needs a sourced answer. No retrieval, weak retrieval, and unread web sources are different outcomes. A model can answer without a source; see Web search.

Declare side effects and risk on tools, configure the applicable approval policy, and follow the action's execution outcome. A prompt asking for approval or an approval_group field alone does not implement every escalation workflow. A completed chat or job does not mean an external action executed.

Desk's audit trail describes the action ledger and its access and retention behavior. Database operators still control the underlying storage; do not confuse API-level restrictions with tamper-proof storage.

Reporting a vulnerability

Follow the security policy for private disclosure, or email contact@meganova.ai.