Skip to content

For Linux servers

Govern agents on the Linux servers you already run

Olivares AI is built to live on the same Linux hosts as the agents it governs. One artifact, installed through the channel you already use, loopback by default. The kernel is local, and this page is explicit about what model inference still sends off the box.

Linux is the house. Coding agents, MCP servers and scheduled jobs land on the same boxes you already patch, log and back up. The missing piece is not another SaaS dashboard: it is a kernel on the host that can answer what those agents can read and write, and stop a governed surface without sitting in the request path.

What this environment gets

  • One artifact next to the agents: a static binary, a container image, or the .deb, .rpm or .apk package when that channel is what you run. The web console is embedded. The published v26.8.0 .apk still carried the systemd unit and no OpenRC unit; Alpine with OpenRC is the later packaging line, not that release asset.
  • Policy at the agent, not at a remote proxy. For Claude Code that is the deny-closed hook; other surfaces attach through connectors, and each connector page says what it can enforce and what it can only observe.
  • An access map built out of band from telemetry and native audit.
  • An estate-wide (or single-agent) stop that is cheap to engage and dual-control to lift. It is a deny gate the governed actuation surfaces consult, not a process killer, and only as wide as the gates you have wired.

Windows is not built. macOS is a supported install target for operators, not the estate the product is designed around.

What this is not

It is not a new distribution, not a hosted Linux fleet, and not affiliated with any distribution or its community.

Related: SRE & sysadmin, home lab, how it works.

Ask Claude

Questions

Does it replace systemd, the package manager, or the distro’s security model?

No. It is a service you install. It does not become PID 1, it does not patch the kernel, and it does not claim to be SELinux or AppArmor.

Can the kernel run with no egress?

Yes for governance data. Hosted models (Claude and other APIs) still reach their provider. Only models you genuinely self-host (vLLM, Ollama) run fully offline.

Try it on your own infrastructure

Olivares AI is open-core (AGPL-3.0) and self-hosted. Deploy it and see what your agents can reach.