HN: HomeBrain – Your private local cloud, agent wired into everything

HomeBrain, a new open-source appliance from developer oalterg, combines Nextcloud, Home Assistant, Vaultwarden, and a local AI agent into a single private cloud device with no subscriptions or cloud accounts. The system, which runs on Ubuntu 24.04 LTS or Raspberry Pi OS, features a master password and recovery phrase, encrypted off-site backups, and a GPU-powered agent that operates via Telegram, with hardware requirements starting at a Raspberry Pi 5 with 8 GB RAM for the HomeCloud variant.

Your private local cloud, agent wired into everything. Nextcloud, Home Assistant, Vaultwarden, and a local AI agent are four products. HomeBrain is the appliance that makes them one: a single password, a backup that restores the whole box, updates that will not brick you, and — on a GPU — an agent that operates it so you never need a shell. No subscriptions. No cloud accounts. No one else's servers. A browser wizard configures it. Nothing leaves the box. Solarized Light, with a dark theme a click away. The dashboard loads nothing from the internet — no CDN, no web fonts, no analytics — so it renders identically on a box with the WAN unplugged. A GPU runs the agent. Without one, the same install is HomeCloud : files, home, vault. ARM or x86 64. One secret, one recovery — dashboard, files, smart home, and backup encryption share a master password. A recovery phrase, shown once, resets the lot. An agent that acts — calendar, lights, files, email. Every token stays on your GPU. Backup that restores the composition — not four separate dumps. Scheduled, encrypted, off-site, one click from the dashboard. It runs itself — pinned versions, downgrades blocked, a factory reset back to the wizard. Health alerts while the box is up. On the LAN, open http://<server-ip . Nextcloud and Vault also speak HTTPS on local ports. With a GPU, Telegram is the daily interface: no VPN, no extra app. Pair it after setup. A public tunnel is optional, if you want the web UIs from outside. Without a GPU, a Pangolin /oalterg/HomeBrain/blob/main/docs/PANGOLIN.md tunnel is how you reach the browser from outside. One domain in, three subdomains out — no port forwarding and no inbound firewall rule. The same screen moves the box to a Cloudflare Tunnel, or reverts every field to what the device shipped with. Dim the living room. Fetch the latest invoice from Nextcloud. Each integration is one row: connect it, test it, revoke it. The agent talks to every service from there. More than one account per service is fine — two homes, two Nextcloud users, several mailboxes. Below the integrations sits the other half: the messenger the agent answers on. Pairing is a code, not a port. Download a model, switch to it, delete the ones you are not using. The previous model stays on disk, so switching back is instant. Weights are fetched once, from the dashboard — never at inference time. A schedule, a drive, and somewhere off-site — the whole backup story is one tab. Archives are encrypted with your master password before they leave the box, so the target never sees your data. WebDAV, SFTP or S3 — or a second HomeBrain, which hands out a dedicated replica account for exactly this. Nextcloud Docker Files, calendars, contacts Home Assistant Docker Smart-home automation Vaultwarden Docker Password manager Bitwarden-compatible MariaDB Docker Nextcloud + Vault database Pangolin Newt Docker Encrypted tunnel optional llama-server systemd Local LLM inference GPU only whisper-server systemd Speech-to-text GPU only OpenClaw systemd AI agent on Telegram GPU only Versions are pinned and updated from the dashboard in one click. | HomeBrain | HomeCloud | | |---|---|---| | CPU | AMD Ryzen 5 / Intel i5 or better | Raspberry Pi 5 | | RAM | 32 GB | 8 GB | | Storage | 512 GB NVMe | SSD | | GPU | AMD Radeon RX 9060 XT 16 GB VRAM | — | | OS | Ubuntu 24.04 LTS | Raspberry Pi OS Trixie 64-bit | Inference is Vulkan via Mesa RADV — no ROCm. Throughput and tuning: BENCHMARKS.md /oalterg/HomeBrain/blob/main/docs/BENCHMARKS.md . Ubuntu 24.04+ / Raspberry Pi OS 64-bit curl -fsSL https://raw.githubusercontent.com/oalterg/HomeBrain/main/install | sudo bash sudo /opt/homebrain/scripts/provision.sh Open http://<server-ip , log in, and initialize. Local network is the default. The wizard shows the generated master password and recovery phrase once . Restoring from backup reuses the old master password instead of generating one. provision.sh prepares the host and starts the manager. It prints a factory password once if you did not pass --factory-pass ; that password opens the wizard and is stored in factory config.txt mode 600 . On AMD GPU hardware a reboot is recommended so kernel parameters take effect; it is not required to open the wizard. Tunnel credentials at provision time are stored on the box and selected in the wizard — they do not skip it. On an already-set-up box the same command repoints the live tunnel without wiping data. Flags and the Pangolin resource map: PANGOLIN.md /oalterg/HomeBrain/blob/main/docs/PANGOLIN.md . | Doc | What's in it | |---|---| | DISASTER RECOVERY.md /oalterg/HomeBrain/blob/main/docs/DISASTER RECOVERY.md PANGOLIN.md /oalterg/HomeBrain/blob/main/docs/PANGOLIN.md ROADMAP.md /oalterg/HomeBrain/blob/main/docs/ROADMAP.md PRODUCT REVIEW 2026-08.md /oalterg/HomeBrain/blob/main/docs/plans/PRODUCT REVIEW 2026-08.md BSD-3-Clause — see LICENSE /oalterg/HomeBrain/blob/main/LICENSE .