deputyOS blog

deputyOS year-one roadmap: where we're heading

Sat, 30 May 2026 00:00:00 GMT

Year one for deputyOS is about going from "engineer-friendly public preview" to "a non-technical user flashes one image and is set up in 10 minutes, on hardware they own, with privacy they can audit." Concrete milestones: domain bootstrap and CDN live (M4a/M4b), egress-whitelist mode, a fourth profile, a desktop launcher notarized on Mac and signed on Windows, and an opt-in cloud backup-restore service that becomes the project's business model.

Give your AI agent your files — safely

Fri, 29 May 2026 00:00:00 GMT

deputyOS lets the agent see your files through three surfaces: VM host-FS passthrough (WSL2 / UTM / qemu), removable drives on bare metal, and SMB/NFS network shares. Every mount lives under /mnt/deputyos// and is policy-gated. AppArmor confines each agent profile to that prefix only. Every mount is reviewable and revocable from `deputyctl mounts`, the wizard, and the PWA — one click to revoke. Credentials never live in the policy file; they live in secrets.env mode 0600.

How deputyOS images are signed (minisign + cosign + SLSA)

Thu, 28 May 2026 00:00:00 GMT

Each deputyOS image is signed with minisign (small, fast, no PKI) at publish time. A cosign signature with SLSA v1 provenance is attached alongside for downstream verifiers. The on-device `deputyctl update` client verifies both signatures, plus the SHA256 of the artefact, before applying any A/B partition swap. The manifest pointing at the artefacts is itself signed. We're working toward SLSA L3 reproducible builds verified by an independent rebuilder.

LFM2 explained: why we picked Liquid AI's models for the air-gapped tier

Wed, 27 May 2026 00:00:00 GMT

LFM2 (Liquid Foundation Model 2) is the model family from Liquid AI we default to in the air-gapped tier. We picked it over Llama-3.2 and Phi-3.5 for three concrete reasons: better per-token quality at the small sizes that fit on Pi-class hardware, an open license that makes baking into an image easy, and a Q4_K_M quantization profile that works well on CPU inference. The 350M, 1.2B, and 2.6B variants map cleanly onto our lean / standard / rich tiers.

Cheap AI on Hetzner: cloud-init recipe walkthrough

Tue, 26 May 2026 00:00:00 GMT

Hetzner doesn't take custom snapshots cheaply, so deputyOS ships a cloud-init recipe instead. Paste the YAML into a CX22 (€4.51/mo) at creation; cloud-init pulls the deputyOS userdata from cdn.deputyos.com on first boot, runs the same wizard. Total setup: under 5 minutes.

Spin up deputyOS on DigitalOcean in 4 minutes

Mon, 25 May 2026 00:00:00 GMT

deputyOS publishes a DigitalOcean snapshot per release. Create a $4 basic droplet from the snapshot, give it 4 minutes to come up, hit http://:8088/wizard, configure profile + provider + channels. You have a working self-hosted AI assistant on a $4/month VPS. The 1-Click Marketplace listing is in progress; the manual path below works today.

Pi 5 vs x86 mini-PC for self-hosted AI: a real-numbers comparison

Sun, 24 May 2026 00:00:00 GMT

Pi 5 is the right starting point — quiet, low-power, $80–120, and good enough for OpenClaw + a cloud LLM. An x86 mini-PC (Beelink / MeLE / NUC class, $150–250) is 2–4x faster for local LLM inference, handles Khoj indexing comfortably, and lets you run rich-tier air-gapped LFM2-2.6B at a usable speed. If you only host channels, Pi 5. If you self-host inference too, mini-PC.

A ChatGPT alternative you run yourself, in 2026

Sat, 23 May 2026 00:00:00 GMT

A self-hosted AI assistant on a Raspberry Pi or mini-PC, running deputyOS, gives you most of what ChatGPT does — chat memory, channel presence, file access — without the SaaS lock-in. You bring your own model provider key (or run LFM2 locally), keep your conversations on your hardware, and pay only for tokens you actually use. Not a perfect replacement for every use case yet (frontier-model coding is still better in the cloud), but for daily-driver assistant tasks it's there.

OpenClaw vs Hermes Agent vs Khoj: three open-source AI agents compared

Fri, 22 May 2026 00:00:00 GMT

OpenClaw is the multi-channel chat-and-skills generalist (Node.js, 23 channels). Hermes is the memory-rich Python agent with a skill system and FTS5-backed recall. Khoj is the personal-knowledge agent that indexes your notes, photos, PDFs, and emails. All three ship as deputyOS profiles and can be swapped on the same hardware via `deputyctl profile switch`. Pick based on what you want first: chat + routing → OpenClaw; long-term memory + skills → Hermes; search + recall over your own content → Khoj.

deputyOS vs Home Assistant: AI assistant vs home automation hub

Thu, 21 May 2026 00:00:00 GMT

Home Assistant is a home automation hub — devices, sensors, scenes, routines. deputyOS is a personal AI assistant — chat, channels, memory, agent skills. They solve different problems and pair well: deputyOS can consume Home Assistant as an MQTT data source via a custom skill; Home Assistant can call deputyOS over HTTP for natural-language intent. Most users want both, on separate boxes.

deputyOS vs Ollama: how they're different and when to use which

Wed, 20 May 2026 00:00:00 GMT

Ollama is a great local LLM runtime — it makes pulling, running, and switching local models trivial. deputyOS is a full personal AI assistant appliance — wizard, channels, memory, hooks, security baseline, signed updates. They're complementary: deputyOS even ships an Ollama provider for users who want to manage their model layer that way. Use Ollama if you want a local model server; use deputyOS if you want a complete always-on AI assistant on hardware you own.

Turn a Raspberry Pi into your personal Telegram AI bot

Tue, 19 May 2026 00:00:00 GMT

Flash deputyOS, boot the Pi, talk to BotFather, paste the token in the wizard. You'll have a Telegram bot that's actually a real personal AI assistant — with persistent memory, multi-channel awareness, and a conversation history that lives on your hardware — in well under 15 minutes. No SaaS, no platform lock-in.

Air-gapped AI on your own hardware with LFM2 and llama.cpp

Sun, 17 May 2026 00:00:00 GMT

Building with AIRGAP=1 produces an image that boots with -net none. A tier-appropriate LFM2 GGUF (350M for lean, 1.2B for standard, 2.6B + Qwen-Coder-1.5B for rich) is baked at /var/lib/deputyos/models/ and served by llama.cpp on 127.0.0.1:8090. apt points at a baked mirror; nftables denies egress except RFC1918, mDNS, and local DNS. Wizard and chat work normally — they just default to the local model.

Run a personal AI assistant on a Raspberry Pi 5 in 10 minutes

Fri, 15 May 2026 00:00:00 GMT

Flash deputyOS-openclaw-rpi5--stable.img.xz to an SD card or NVMe with Raspberry Pi Imager. Boot the Pi. Within 60 seconds the wizard publishes itself at http://deputyos.local. Scan the QR shown on HDMI (or open the URL on your phone), pick the OpenClaw profile, paste an OpenRouter or OpenAI key, link your Telegram bot. Total hands-on time: about 10 minutes.

Why deputyOS does zero first-boot installs

Thu, 14 May 2026 00:00:00 GMT

deputyOS images contain every runtime, native module, model file, signature DB, and configuration template needed to run the chosen agent. First boot only does configuration — no apt, no npm, no pip, no git clone. This makes installation reliable for non-technical users, makes the image reproducible byte-for-byte by independent builders, and makes air-gapped deployment a flag rather than a redesign.

Why local-first AI matters now (and what it costs you)

Tue, 12 May 2026 00:00:00 GMT

Local-first AI runs the model — or at least the agent and its memory — on hardware you own. The benefits are concrete: no data lock-in, predictable cost, no SaaS-provider whim, and conversation privacy that survives policy changes. The costs are honest: slower inference on consumer hardware, you maintain the box, and not every cutting-edge capability is available offline yet. We think the trade is worth it for most personal assistants, and getting better every quarter.

Try deputyOS on your laptop without buying anything

Mon, 11 May 2026 00:00:00 GMT

deputyOS runs in a VM on your existing laptop or desktop. macOS users install UTM; Windows users run `wsl --install -d deputyos`; Linux users use the qemu/KVM that's already on most distros. The same image you'd flash to a Pi runs as a local VM. About 2 minutes once you have the VM tool in place. When you outgrow it, the same image runs on a Pi or a cloud VPS — your conversation memory and config back up cleanly.

Why we built deputyOS: a personal AI you actually own

Sun, 10 May 2026 00:00:00 GMT

We built deputyOS because the existing path to running an open-source personal AI assistant — install Node, install Python, fight native modules, configure systemd, lock down a network, schedule backups — loses more users than the agents are worth. deputyOS bakes everything into one signed appliance image. Flash, boot, scan a QR, set a key. Local-first by default. Air-gapped tier optional. Apache-2.0.

Hello, public preview (v0.3.0)

Fri, 08 May 2026 00:00:00 GMT

deputyOS — the batteries-included Linux appliance image for personal AI assistants — is publicly previewable as v0.3.0. New in this release: a drive-mounting subsystem (host FS, removable, SMB/NFS), an air-gapped tier that bakes LFM2 via llama.cpp so the appliance works with no network, and three public sites at www / api / cdn .deputyos.com.