Assets & Inventory

Hypervisor hosts

Hypervisor hosts

If a machine in your estate is a hypervisor — a Proxmox VE node, a Hyper-V server, a Linux box running KVM through libvirt — the reporting agent installed on it does one more thing: it reports the guests that host is running, so the virtual machines and containers living on it appear on your map without an agent inside each one, connected to their host by the runs-on link that makes blast radius mean something.

There is nothing to set up for this. You install the same agent, with the same one-liner, on the hypervisor host exactly as you would on any other server — no API token, no extra credential, no separate collector to enable. The agent detects for itself that it is running on a hypervisor and starts including the guest list in its ordinary reports; the installer prints what it detected (“Detected: Proxmox VE — this host’s QEMU VMs and LXC containers will be inventoried”) so you know before the first report lands. Detection is re-checked on every report, so a machine that becomes a hypervisor later — you enable the Hyper-V role, you install Proxmox — simply starts reporting its guests on the next check-in, with nothing to re-install.

It works on:

  • Proxmox VE — QEMU virtual machines and LXC containers, read locally on the node itself.
  • Hyper-V — through the Windows agent already on the host.
  • libvirt/KVM — a generic Linux host running virtual machines under libvirt.

The guest list is still the local machine describing itself — the host asking its own hypervisor what it is running, the same way a Docker host reports its own containers. The agent never scans your network and never calls a remote hypervisor API.

VMware is the honest exception: ESXi hosts cannot run agents — the platform does not allow installing third-party software on the hypervisor itself — and vCenter support is planned as a server-side connection, configured in lazyit rather than installed on a host. Until that ships, VMware guests appear the way everything else did before agents: run the agent inside the guests you care about, or add them by hand.

One more platform is worth naming: on XCP-ng / XenServer the agent detects the hypervisor and says so, but does not collect its guests yet — an honest “detected, not inventoried” beats a half-working list.

What appears, and where

Each guest the host reports lands in the Pending review tray on the Servers view — the same tray, the same rules as every discovery — grouped under its hypervisor host, and connected to it on the map by a runs-on link once confirmed. Nothing enters your live inventory until you confirm it, one at a time or in bulk.

  • A QEMU, Hyper-V or libvirt guest is proposed as a virtual machine node; a Proxmox LXC container is proposed as a container node.
  • A guest is matched by the platform’s own stable identifier (the Proxmox VMID, the Hyper-V VM GUID, the libvirt domain UUID), so renaming a VM does not create a duplicate.
  • Like containers, guests default to not being tracked as assets when you confirm them — a homelab-sized host can carry dozens — and the toggle is right there if a VM is something you track.
  • A guest that disappears from the host’s report — deleted, or migrated away — has its node marked offline. It is never removed behind your back; Discard stays your call, and the same guest returning brings the node back online.
  • On a Proxmox cluster, each node reports only its own guests, so a cluster where every node carries the agent covers the whole estate with no overlap.

A busy host can propose a lot at once. Guests enter through the same discovery limits as everything else, so a host with hundreds of VMs fills the tray gradually over an hour or two on its first report rather than all in one burst — that is the flood protection working, not guests being lost.

One machine, one node

The obvious worry: if a VM runs its own agent and its hypervisor reports it, do you get two nodes for one machine? No — they converge automatically. The hypervisor knows each guest’s firmware identity (its SMBIOS UUID) and its network-card addresses; the agent inside the guest reports the same facts from the other side. When both match — the identity corroborated by a network-card address, never on a single signal — lazyit folds the host’s view into the guest’s own node: the in-guest agent’s record wins (it has the real inventory — the software, the disks, the machine identity), and the runs-on link lands on it. You end up with one node that both knows what it runs and where it runs.

When the corroboration is not there — cloned VMs really do ship duplicate firmware identities — the two rows are surfaced as a possible duplicate for you to check and merge, never merged for you.

A guest with no agent of its own is simply represented by the host’s view: a confirmable, trackable node that goes offline when the guest does. Install an agent inside it later and the two converge the same way.

When a guest stays as two rows

Sometimes the automatic merge cannot happen, and lazyit now tells you instead of staying quiet about it. The automatic merge needs the guest’s firmware UUID, reported by the agent inside the guest. When that agent reports no firmware UUID, lazyit falls back to matching the network card — and a network card alone is a hint, never a merge. You will see a possible duplicate notification saying a network card matches but there is no UUID to confirm it.

The notification tells you which of the three causes you are looking at, because the fix is different for each:

  • A container (LXC). Containers have no firmware UUID of their own — there is nothing to read, at any privilege level. This pair can only be closed by hand, and that is fine: merge once and it stays merged.
  • An agent running unprivileged. On Linux the firmware UUID is readable only by root, so an agent running as an ordinary user simply omits it. Run the agent as root (or Administrator on Windows) and the two rows converge on their own from the next report — no merge needed.
  • A VM whose firmware Windows cannot read. Some virtual machines present their firmware in a form Windows cannot locate, so Windows reports no SMBIOS UUID at all. This is common on Proxmox VE and only for Windows guests: Proxmox freezes a Windows VM’s virtual hardware version when you create it, so a VM created on PVE 8.1 or 8.2 keeps that hardware forever while the Linux VMs beside it are unaffected. Raise the VM’s Machine version (VM → Hardware → Machine) to a current one, or add -machine smbios-entry-point-type=32 to its args, then reboot the guest — the two rows converge on their own from the next report onwards.

In every case you can also just merge them once, from the tray: Merge into… folds the hypervisor’s row into the guest’s own node, keeping everything you set.

Upgrading to this version? The first report from each host will raise these notifications for pairs that were already silently split — you are seeing existing duplicates surface, not new ones being created. Rows that already forked do not merge themselves. You get one notification per (guest, network card), not one per report, so each pair rings once and then stays quiet — but if your estate runs many LXC containers with their own agents, or many unprivileged agents, expect that first wave to be large, because every one of those pairs is genuinely un-mergeable automatically.

Cluster migrations

When a VM migrates between Proxmox cluster nodes, node A stops reporting it and node B starts — so its old node goes offline under A and a new pending proposal appears under B. lazyit notices the two look like the same machine and shows the duplicate-suspicion hint in the merge dialog; Merge into… folds them into one, keeping everything you set. This is deliberately a one-click suggestion rather than an automatic merge — a wrong merge is expensive, a hint is cheap. A guest running its own agent skips all of this: its one canonical node just gets its runs-on link re-pointed to the new host.

Turning it off

Guest inventory follows the same three controls as every other collector, and the off switch always wins over the on one:

  • At install time — add --no-hypervisor to the Linux install command, or -NoHypervisor on Windows. This writes the veto below into the host’s config for you. The create-agent wizard carries the same veto as a checkbox: Advanced options → Don’t inventory this host’s guests appends the right flag to the command it generates.
  • On the host — set LAZYIT_COLLECT_HYPERVISOR=false in the agent’s own config file (/etc/lazyit-agent/config on Linux, C:\ProgramData\lazyit-agent\config on Windows). As with every local setting, this wins over anything set in lazyit and survives agent upgrades.
  • Centrally — the Hypervisor guests switch in Settings → Reporting agents, which turns it off for every agent in the estate on their next check-in. It is on by default: detection already gates collection, so on a host that is not a hypervisor the switch changes nothing.

Turning it off stops future guest reports; the guest nodes you already confirmed stay where they are, going offline as their reports stop, and Discard remains yours.

The agent reads the guest list — names, identities, state, sizing — never the guests’ contents. It does not look inside a guest’s disks, does not need an agent or any tooling installed in the guests, and sends no metrics. It is inventory, in the same narrow sense as everything else the agent does.