Skip to content

How it works

A kernel, the agents around it, and the connectors between them.

Olivares AI is one self-hosted artifact. The kernel is the engine that stores policy, the access map and the audit ledger. Agents — coding sessions, MCP servers, scheduled jobs — attach through connectors. The picture on this page is the out-of-band kernel: it does not sit on the model-call hot path unless you opt into the inline /v1/messages enforcement point.

The pieces

Four names, one binary

Kernel

The engine inside the single artifact: policy, the read/write access map, the append-only hash-chained audit ledger with signed checkpoints, and the optional deny-closed enforcement points you wire in — in current development the Claude Code hook, an opt-in inline /v1/messages proxy that stays unmounted until you provision it, an MCP tools/call gate and an A2A delegation gate. It is not a Linux kernel and not an inference runtime.

Agents

Sessions and workloads you already run — Claude Code, Codex, Grok Build, MCP servers, scheduled jobs. They keep their own model calls. The kernel sees identity, what each agent reached, and whether that was permitted.

Connectors

An Apache-2.0 SDK. Sources (runtimes, identity, stores, audit trails, OpenTelemetry) push edges in. Outputs (console, CLI, API, SIEM) take findings out. Collectors run in your infrastructure, read the systems you already operate and push observations to the kernel; they do not sit in the production data path and do not intercept model-call traffic.

Integration

The same API under the embedded console, the CLI and the Terraform provider. Policy and map stay in infrastructure you control. Hosted models still reach their provider; only models you genuinely self-host run offline.

Agents attach to the kernel. Policy and evidence travel on those edges — not prompts, not payloads. KERNEL policy · map · ledger deny-closed gates (opt-in) Coding session Claude Code · Codex · Grok MCP server tools the session can call Scheduled job unattended agent Human operator console · CLI · approvals POLICY AND EVIDENCE — NOT PROMPTS

Agents attach to the kernel. Policy and evidence travel on those edges — not prompts, not payloads.

Connectors bring native audit and telemetry in, and send governed findings out. The kernel does not replace your gateway, tracer or IdP. SOURCES IN OUTPUTS OUT Runtimes & MCP Identity & stores OpenTelemetry & audit Native trails KERNEL edges, not payloads Apache-2.0 connector SDK Console & CLI Stable API SIEM / your rail Terraform provider

Connectors bring native audit and telemetry in, and send governed findings out. The kernel does not replace your gateway, tracer or IdP.

Olivares AI is a pre-1.0 preview. This page describes how the product is designed to run. It is not certified. Air-gap applies to the kernel and its data, not to hosted-model inference.

See it on your own infrastructure

Install the artifact, attach one agent surface, and read the access map. Architecture and solutions sit next to this page.