Aug 2026 · self-hosting, Raspberry Pi, agent deployment
After the dashboard post, the next question was: what happens when the Win Mini sleeps? So I gave the fleet a second home: a Raspberry Pi 3B+ with one gigabyte of RAM, named jacinto.
The brain is thin: the model lives on the API, not the hardware. Same Hermes, same
llm_client.py, and nobody on Discord can tell which machine answers. A
local-only brain could never fit in 1 GB.
A host on the LAN accepted the user pi without being the Pi: verify
the MAC (b8:27:eb:…) before trusting anything. And the modern Pi OS
ignores the country= line in wpa_supplicant.conf: Wi-Fi
stays rfkill-blocked until raspi-config nonint do_wifi_country.
The install kept dying with the SSH tunnel, so it runs under
systemd-run, immune to my connection dropping:
ssh jacinto 'systemd-run --user --unit=hermes-install --collect \ bash -c "curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash"'
Cloud provider only (a quantized 8B does not boot in 1 GB), and its own Discord bot token: one token is one connection. The bot is my ops console. I typed "deploy alienhabits" and it did: wrote the unit, documented it in its own skill, scrubbed the password from the chat. Vaultwarden, the sync, the exporter binds — all managed from its chat.
Twelve repos cloned over SSH (by rule, bidirectional). Syncthing keeps the vault
(8,617 files) and backups on both machines. node_exporter feeds a
jacinto-health dashboard (fourteen panels), and Prometheus plus the
Pushgateway answer from the Pi when the Win Mini sleeps. One-gigabyte budget:
zram swap, Bluetooth off, journald capped.
The brain runs wherever it is installed, the data lives in two places, and the second home cost the price of a microSD card. The API won, again: this time it bought me a second house.