Blog
A ChatGPT alternative you run yourself, in 2026
Why and how to replace ChatGPT with a self-hosted AI assistant on hardware you own — without giving up the things that make ChatGPT useful.
“What’s a ChatGPT alternative I can run myself?” is one of the most-searched questions about local AI. Most answers point at Ollama or LM Studio. Those are great if you want a local model — but they’re not a ChatGPT experience replacement. ChatGPT is more than a model: it’s a chat UI, memory, files, voice, plugins. Replacing it means replacing those.
deputyOS is, by design, that replacement.
What you actually use ChatGPT for
Before picking an alternative, name the use cases. For most people:
- Conversational chat. Ask things, get answers, follow up.
- Memory across sessions. “Remember that I prefer SI units” → it does.
- File / image upload. “Summarise this PDF” / “what’s in this photo.”
- Channel presence. ChatGPT in the iPhone share sheet; sometimes via mobile app keyboard.
- Voice. “Hey ChatGPT, what’s the weather in Lisbon.”
- Plugins / tools. Web search, code interpreter, image generation.
A real alternative needs to cover most of these, not just point 1.
How deputyOS maps to those needs
| Use case | deputyOS coverage |
|---|---|
| Conversational chat | /chat web UI on the device; PWA at /app |
| Memory across sessions | Per-profile persistent memory; OpenClaw simple log, Hermes FTS5, Khoj vector index |
| File / image upload | Drive-mounting subsystem for local files; Magika type-sniffs uploads |
| Channel presence | Telegram, Slack, Discord, Signal, Matrix, IRC, WhatsApp, Email — pick in the wizard |
| Voice | whisper.cpp wake-word + Piper TTS, on-device on rpi5 |
| Plugins / tools | Skills system per profile; web search via skill; no image-gen yet (on the roadmap) |
What you give up vs ChatGPT
Honesty matters. The trade-offs are real.
Frontier-model output quality is still cloud-better
For deep code generation, complex reasoning, long-context multi-doc analysis — GPT-class models in the cloud are still ahead. deputyOS lets you point at OpenAI, Anthropic, OpenRouter, or Bedrock for these (bring your own key), so you can have the same model quality, just routed through your own agent + memory layer. The pure “local-only LLM” path (LFM2-1.2B / 2.6B) is great for short-form chat, intent routing, summarisation — but not yet at frontier-model parity for hard tasks.
Mobile app polish
ChatGPT’s mobile app is excellent. The deputyOS PWA at /app is good — fully installable, works offline-once-cached — but it’s not pixel-perfect-polished the way a native app is. Most users acclimatise within a day.
Image generation
Not yet baked. The skill system can call out to a separate image-gen service if you set one up, but we don’t ship one on-device. Frontier image gen needs a GPU you probably don’t have on a Pi.
”What everyone else is using”
If your social signal value is “I’m a ChatGPT power user,” self-hosting doesn’t carry that signal. (This is a real thing for some users.)
What you gain vs ChatGPT
Privacy that’s structural
The conversation data lives on your hardware. Future ToS changes can’t reach it. The provider you pick sees the queries (because they answer them); they don’t see your memory, your channels, your skills, or your file access.
Cost predictability
You spend tokens at whichever provider’s rate you chose. Cost guardrails auto-pause at the cap. There’s no per-seat subscription that drifts upward annually.
Channel uniformity
ChatGPT lives in ChatGPT. Your deputyOS assistant is on Telegram, Slack, Discord, IRC, Matrix, Email — all routing to the same memory. One assistant, many surfaces.
Air-gapped option
If you want truly offline, build with AIRGAP=1. The image bakes LFM2, locks egress at nftables, and you have a working chat with zero outbound network. Air-gapped AI walkthrough.
No deprecation pain
When a cloud LLM gets deprecated, your historical conversations with it become slightly inconsistent (the new model “remembers” differently). With deputyOS, you pin model versions per provider; you choose when to migrate.
The hardware math
You need a place for the agent to run. Options:
- Raspberry Pi 5 8 GB: ~$80, runs OpenClaw or Hermes great, ~5 W idle. Setup: 10-minute walkthrough.
- Pi 5 16 GB: ~$120, comfortable for Khoj indexing + air-gapped rich tier.
- x86 mini-PC (Beelink/MeLE): $150–250, fastest experience, ~25 W under load.
- $5/mo cloud VPS (Hetzner / DigitalOcean / Linode): zero hardware investment, see DigitalOcean walkthrough.
- Desktop launcher (Win/Mac/Linux): runs in WSL2/UTM/qemu on a machine you already own. Lower friction; needs the machine on.
Power costs and Telegram/Slack token costs aside, you can absolutely run a Pi 5 + an OpenRouter key for $5/month total and have an assistant that’s qualitatively better than ChatGPT free.
How to switch in stages
You don’t have to commit cold-turkey. Reasonable order:
Stage 1: Set up deputyOS in parallel
Get a Pi 5 (or use WSL2 / cloud VPS). Wizard, OpenClaw, an OpenRouter key. Wire up Telegram. Use it alongside ChatGPT for a week. See if it covers your daily needs.
Stage 2: Mount your files
Mount the folder where you keep notes / docs / receipts. Now your assistant can read them. The ChatGPT replacement bar is starting to look more reachable.
Stage 3: Try Khoj
Switch profile to Khoj (deputyctl profile switch khoj). Index your Obsidian / Notes folder. Now you have search-and-chat over your own writing.
Stage 4: Cancel ChatGPT Plus
For most users, this is when it stops being the daily driver. Your deputyOS assistant has memory, channel presence, file access, and your context. ChatGPT was a habit; now you have your own habit.
Stage 5: Try air-gapped
Optional. Build with AIRGAP=1. Now even the model is local. Slower for hard queries; the option exists when you need it.
When deputyOS is not the right call
- You write production code all day with frontier-model assistance — keep using ChatGPT or Claude for that workflow, point deputyOS at the same provider for non-coding tasks.
- You don’t have any spare hardware and don’t want a cloud VPS — the entry cost is real.
- You need a Slack-app-store-style “company assistant” with audit trails and admin controls — that’s enterprise SaaS territory.
The deeper point
ChatGPT is a wonderful product. deputyOS isn’t trying to beat it on shine. It’s trying to beat it on structural ownership: your conversations, your hardware, your provider choice, your memory format, your channel set, your update cadence.
For a personal assistant, structural ownership compounds. A year from now, an deputyOS user has a year of conversation memory they own. A ChatGPT user has a year of conversation memory locked in a SaaS account that may or may not still exist in its current form.
The picker is where you start. The 10-minute Pi 5 setup is where you get there fast.
Related posts
deputyOS vs Ollama: how they're different and when to use which
Both run AI locally. They solve different problems. Here's a real comparison so you can pick the right one for your use case.
Air-gapped AI on your own hardware with LFM2 and llama.cpp
Build an deputyOS image that bakes LFM2 via llama.cpp, locks egress at nftables, and runs a full personal AI assistant with zero internet.