Skip to content
A personal experiment · alpha · code not public yet · no offer, no support Project status →

Action catalogs

Vaicoa knows actions only as entries in a signed catalog. Nobody, human or agent, can run anything that is not in it. Each entry says:

  • the parameters, as a JSON Schema the core validates;
  • the risk class: safe, routine, risky (and never in policy);
  • the constraints, CEL expressions over the parameters and the state a module supplies;
  • the required level of the proposer and the approval rule (the quorum: how many people, which roles, which level, which channels, how long the approval is valid);
  • the module that supplies state and the plan, and the executor that runs it (on the node or at the core);
  • the display template: how the approval looks on every device;
  • whether it is destructive (the approver then also picks or types the target).

The format is in the reference.

Version 2026.11.9 has 74 actions: 15 safe, 36 routine, 23 risky. They cover Podman environments, hosts (OS updates, reboots, isolation), k3s, Kubernetes workloads and volumes, Flux, CloudNativePG backups and restores, Cluster API, Hetzner Cloud servers and firewalls, Forgejo (git.change, git.tag), incidents and alerts, access grants, certificates and secrets, restore drills and game days. On top of them come five playbooks: patch.window, release.rollout, incident.first_response, restore.drill and gameday.run (operations processes).

The catalog is signed with the release key and pinned by the core and every node. A node refuses an op whose action is not in its pinned catalog.

Five actions only read: metrics.query, logs.query, traces.get, kube.get, kube.logs. They need no approval and run at once, but they are narrow: a read budget per run, redaction of secrets before the answer leaves, and in a production environment nothing at all unless you switch reads on for it (agent-read). The audit records that a read happened and its size, not its content.

An agent session sees only the actions its trigger allows (mayPropose), as tools of vaicoa-run. Name, description and input schema come only from the verified signed catalog, never from text an agent or a third party can influence; if the catalog’s signature does not verify, the MCP server does not start.

Your own actions live in an internal layer org.<name> next to the built-in catalog, signed by your team’s owners and pinned in the roster. A layer can add actions and playbooks; it cannot change a built-in one. A layer action appears in the approval as “from catalog org.…”. The user manual’s chapter 18 walks through one (notify.send), by hand and with a coding agent.

If a provider offers an MCP server, vcx catalog import-mcp drafts a layer from it. Only reading tools are taken over, as safe reads with a budget and redaction; tools for arbitrary commands or SQL are always left out. The draft is a proposal: it becomes binding only with your owners’ signatures. If the provider later changes a tool, the core marks the action unavailable (MCP DRIFT) and refuses it until you sign a new layer.