Architecture · How it works
Olivares AI runs where your agents run. It deploys as a single self-hosted artifact, discovers passively without sitting in the request path, and the control plane is built to run fully air-gapped. This is how the pieces fit together.
How it runs
Install the engine as a single self-hosted artifact — a binary, a container or on Kubernetes. The web console comes embedded; there is no separate front-end to stand up.
Place collectors inside your own infrastructure. They passively ingest the OpenTelemetry your agents emit and read the native audit trails of the systems they touch — read-first, minimum-privilege, push-only.
The platform normalises those signals into a model of agents, sessions, models, MCP servers, resources and identities — and the access edges between them, read or read/write, each with an honest confidence level.
On that map, see the gap between what an agent is permitted to access and what it was observed accessing, enforce policy, and route findings to the tools you already run — all via CLI, API and the web console.
How it is built
The collectors run inside your own infrastructure and capture edges and relationships — not payloads. The control plane is self-hosted (a managed option is on the roadmap). Self-hosted means the data does not leave your environment.
The engine is CLI-first and exposes its full functionality through a stable API; the web console is a layer over that same API, shipped embedded. You deploy one thing, not a fleet of services.
It listens to telemetry your agents already emit and reads native audit trails, instead of intercepting traffic. A stable, Apache-2.0 connector SDK covers a growing set of sources — models and providers, runtimes, data stores, identity and secrets, observability, and output destinations.
It ingests OTLP telemetry and interoperates with the observability stack you already run — exposing standard metrics and health endpoints and exporting to common SIEM formats. It slots into your pipelines instead of replacing them.
The control plane is designed to run fully air-gapped: governance and observation data stays local, with zero egress, licensing validated offline, and an install bundle you can verify without network access. Hosted models (Claude and other API providers) still reach their own API; only genuinely self-hostable models run fully offline. Deployable in the most restricted and regulated environments.
Read-first and asymmetric-low-risk. Collectors are push-only with no inbound listener, traffic to the core is mutually authenticated with TLS, only the minimum data is kept, the audit trail is append-only and tamper-evident, and multi-tenant isolation is built into the model from day one.
Deployment
The same product, deployed the way your environment needs it.
The engine and the embedded web console as one self-hosted artifact — a binary or container, or installed on Kubernetes — for a single node.
Collectors across many hosts feeding a central control plane with an external database, for multi-host scale and multi-tenant operation.
The control plane and collectors run locally with zero egress and offline licensing, for disconnected, regulated or classified networks. Agents on hosted models still need outbound access to their provider; agents on self-hostable models run fully offline.
The control plane hosted by us — while the collectors always remain inside your own infrastructure.
Olivares AI is in active development. This page describes how the product is designed to run; the access map and least-privilege drift are what the platform is built to surface, framed honestly until the product ships.
Self-host the complete product on your own infrastructure, or talk to us about a distributed or air-gapped deployment.