injection_guardHostile content is stripped
Instructions hidden in a document, an HTTP response or a database row are removed at the boundary, before the model reads them.
Features
Agents, knowledge, data access and operations on a single engine that runs on your GPUs, under your identity, with every step audited.
How a request flies
Every request passes the same stages, in order. Observe loops back to plan until the result is verified against real tool output.
Agents & skills
An agent is a role, a skill set, credentials and policy. Business owners assemble them; skills are packaged and shared as files.
Knowledge: RAG & OCR
Contracts, invoices, drawings and scans become searchable, and every answer carries its sources.
Data & integrations
AOS reaches the systems where answers live. Credentials are leased per user, so row-level security and audit stay intact.
Platform & operations
Models are pluggable and pinned. Nothing is repriced or deprecated underneath you.
Security
AOS assumes the model is untrusted and the data is hostile. Both assumptions are enforced in code.
injection_guardInstructions hidden in a document, an HTTP response or a database row are removed at the boundary, before the model reads them.
secret_scrubberCredentials in error traces and config dumps are scrubbed from text before it reaches any model, local or public.
env_vaultEnvironment settings live in a sealed, encrypted vault rather than an unprotected file on disk.
credentialsEncrypted credentials are leased per user, so row-level security still applies. AOS fails closed and never falls back to a shared account.