Launch, attach and govern the agents you drive

You already drive coding agents. Olivares AI is the kernel those sessions attach to: deny-closed policy at the hook, launch and attach from the console, budgets, a kill switch, session recording. Inference still goes to the provider; governance stays on your side.

The daily loop is the point: start a session with a permission mode and a workspace, attach to its I/O, stop it, resume it. Two sessions should not hold the same piece of work — that is what work items and fenced leases are for.

Connectors keep Claude Code, Codex and Grok Build on the same kernel. They are not three equal in-path gates: Claude Code is the deepest in-path surface, and what each of the others can enforce versus what it can only observe is stated on the connector pages — Grok Build, for one, enforces at PreToolUse and observes the rest — not flattened into one claim.

Beside the tools you already use

It does not replace git, CI or the agent CLI. It sits beside them. Policy that cannot run at the hook is not presented as in-path.

Questions

Does this make Claude self-hosted?

No. Anthropic publishes no weights. The kernel and the evidence stay on your side of the wire; the model call does not.

Is the cockpit finished?

Launch, attach, stop, resume, budgets and session recording exist as API and console surfaces. Claude Code is the deepest in-path surface, through the deny-closed hook; Codex and Grok Build are first-class command surfaces on the same kernel. The depth of the session console is still being aligned with the architecture (pre-1.0). Where a page marks live I/O as preview, treat it as preview.

Beside the tools you already use

Questions

Try it on your own infrastructure