A self-installing AI-agent appliance. Ollama runs as a loopback-only Windows service, gemma4:latest is the local model, and a modern computer-use chat client lets you talk to Gemma 4 and let it drive the machine.
Everything is local — the model, the API, and the agent. Nothing leaves the loopback interface.
Automatic (delayed) LocalSystem service on a first-party OllamaServiceHost.exe that supervises ollama serve with 5s/15s/60s restart backoff. No NSSM.
gemma4:latest served over 127.0.0.1:11434. Thin-ISO pulls it on first boot with progress shown in the chat client; or embed it in the image.
Modern .NET 8 WPF client, bottom-right at logon. Chat with Gemma 4 and it proposes actions: run PowerShell/cmd, open apps, files, keystrokes, clipboard, screenshots.
Bound to 127.0.0.1, no inbound firewall rule, local-only CORS origins. The installer asserts loopback binding at first start.
Defender, Windows Update, Secure Boot support, servicing, networking, PowerShell, recovery — all intact. No security controls are disabled.
Agent ships via sources\$OEM$ so install.wim stays pristine and Windows Setup installs fully unattended, straight past the product key.
Six stages from clean checkout to a validated ISO. Expand any step to see what it actually does and the exact command that runs.
Expected outputs for every gate — checksums, bootable-ISO structure, and the Gen2 UEFI boot test. Failures highlight in red with the exact reason and remediation. Toggle a row's simulate fail to see how a regression surfaces.