Updated 2026-09-28, after the Ubuntu agent computer was removed and Omarchy provisioning was automated.
Clusters and machines
| Cluster | Where | Contents |
|---|---|---|
32 devserver |
<old-devserver-ip> |
Old container-era cluster (the testing computer was deleted) |
33 devserver |
Hetzner dedicated, Helsinki (<hetzner-ip>) |
KubeVirt; the hand-built Omarchy (id 11, ns omarchy-spike); omarchy-auto, the first computer Canine built by itself; the user's own agent-desktop VM (don't touch). desk-2 and the golden image are deleted. |
34 cluster |
AWS m7i-flex.xlarge, us-east-1 (<aws-node-ip>, no Elastic IP yet) |
k3s v1.36, KubeVirt + CDI, the hand-built Omarchy (id 12, ns omarchy-spike). Golden image deleted. |
Study material: this lesson, in the ai-lessons repo under books/canine/agent-computers/.
Files outside the repo: ~/Downloads/aws-devserver-kubeconfig.yml (cluster 34 kubeconfig) and
~/Downloads/devserver-keypair.pem (EC2 SSH key). The spike's Omarchy SSH key, login password, VM spec and
unattended-install config are in labs/.secrets/ (git-ignored, local only). Computers Canine builds keep their own
password and SSH key on their database record.
What changed on 2026-09-28
- Removed: the Ubuntu/XFCE image (
provision.sh), the golden-image build (AgentComputer::Image,BuildImageJob), the X11 computer server and its Control tab and/apiproxy route. See the notes at the top of The golden image pipeline and The computer server. - Added: "New computer" now builds an Omarchy VM end to end. See Automating Omarchy.
- Kept: the proxy (Selkies + VNC), the connect page, stats, and the KubeVirt installer (minus the image build).
Deferred work
- Lifecycle: Stop and Start work (a stopped computer keeps its disk and frees its memory and CPU; start is ~30 s to a streaming desktop). Still to do: stop idle computers automatically, start on connect.
- Agent control on Wayland: built (in review): a computer-use server in each VM plus MCP tools, with its own
book, Computer Use. Still to do: API tokens on the proxy and
/api/v1/agent_computers. - Share the ISO: each computer imports its own 6 GB copy; import once per cluster and clone it instead.
- Latency tuning: Selkies encoder settings, CPU measurement, instance type, maybe WebRTC.
- Ops: Elastic IP for cluster 34.
- Disk stats: Omarchy has no QEMU guest agent, so the overview can't show disk usage. Installing it needs a full
pacman -Syu, which provisioning deliberately avoids.