Nguyen LNP AI / Service
Self-hosted AI systems
Your own AI assistant, on a box you control
I set up personal AI systems on a VPS, a mini PC or an Android phone, then keep them running: updates, cleanup and the occasional move when the tools change. You talk to it on Telegram. It remembers what it learns in a wiki you own, and it asks before anything leaves your hands.
Who it's for
People and small businesses who want an assistant that actually works for them, not another chat tab. Usually that means one of these:
A second brain
Notes, saved links and decisions you can ask questions of later, kept in a wiki you can read without the AI.
A morning brief
A daily summary of news, tasks and your calendar, sent to the same chat you use for everything else.
Project admin
Planning, tracking and reporting on work, with the agent doing the updates and you doing the decisions.
It's less of a fit if you need everything to run with no internet connection: the agents use Claude, which runs on Anthropic's servers. More on that below.
What "self-hosted" means here
The agents, their memory, their tools, their logs and your secrets live on hardware you control. The model doing the thinking is Claude Opus 5.5, which runs on Anthropic's side, so the system needs an Anthropic account. Jobs that should stay on the machine, such as image, video and voice generation, run on local models where the hardware can handle it. My own media box does that on a single 8 GB graphics card.
One design rule carries over from my own setup: the important parts keep working without the AI. The wiki is plain pages, synced through GitHub and never used for secrets. My own project board reads its cards straight from the wiki, so it still works when Claude isn't running.
Where it can run
| Hardware | Good for | Worth knowing |
|---|---|---|
| VPS | An assistant that is always on and reachable from anywhere. | Usually no graphics card, so media generation runs elsewhere. |
| Mini PC | An always-on box at home or in the office that you physically own. | This is the always-on machine in my own setup. |
| Android phone (Termux) | An agent that travels with you and can work with what's on the phone. | In my own setup the phone is one of four devices. |
What you hand over, what you get back
| You hand over | You get back |
|---|---|
| The device or VPS it should live on, and admin access for the setup. | Claude Code installed and running on it. |
| The jobs you want it to take on, in plain words. | A Telegram chat (or one topic per project) where you give it work and it reports back. |
| The accounts and tools it should reach, through a safe login route we agree on. | Tool connections, with secrets kept out of the wiki and out of chat history. |
| Who approves actions that leave your hands. | An approval step: it asks on Telegram before anything that saves, sends, submits, deletes or touches a password, and logs every answer. |
| Your notes and documents, if you want them in its memory. | A shared wiki the agents read before they start and write back to when they finish, plus a runbook covering restarts, updates and restores. |
How an engagement runs
- Brief. Send what you want the assistant to do and where it should run, on Telegram, WhatsApp or email.
- Written plan, then your sign-off. I write down the hardware, the tools it will reach, what it may do on its own and what must wait for you. Nothing gets installed until you approve it.
- Setup. Install, connect the tools, create the wiki and the Telegram chat, and wire up the approval step.
- Test the gate. We run a real job that ends in a send or a delete, and check that it stops and asks you first.
- Handover. You get the runbook and a walkthrough. The approval step stays on.
- Maintenance, if you want it. Updates, cleanup, and moving the system when a tool is replaced.
Runs on its own
Reading, searching, typing, navigating, drafting and writing to its own wiki. Nothing has left your hands yet.
Waits for your OK
Saving over records, sending, submitting, publishing, paying, deleting, and anything that touches a password.
Worked example: moving my own four devices to Claude
No client names here, so this is my own system. It is the "maintenance and moving" part of the service, done on myself first. For most of 2026 my agents ran on OpenAI models through Hermes Agent, ending on GPT-6. Then Claude Opus 5.5 got so good that I moved every device back to Claude.
- Hermes agents on GPT-5.5, GPT-5.6 and GPT-6 ran my publishing pipeline and coding work from June to September.
- A Claude Code hub on the PC starts taking work from a Telegram group, alongside the old agents.
- Claude only. No Gemini, ChatGPT or DeepSeek in the loop. One model across every device.
- Last Hermes agent retired. Its projects moved to Claude Code.
What it looks like now: a Windows PC, an always-on mini PC, a ROG Ally handheld and an Android phone in one Telegram group, with a topic per project and one bot per device, a shared handoff board so devices pass work to each other, and a wiki of about 800 linked pages that every agent reads before it starts. A client setup is the same pattern at a smaller scale, for example one device and one chat.
Guides worth reading first
The pieces a small team needs, and in which order.
Self-hosted AI backup and recovery checklistHow to restore the whole agent stack, not just the files.
MCP server security checklist for self-hosted AIWhat to lock down before an agent gets access to your tools.
Tell me what you want your assistant to do
Say where it should run and which tools it needs to reach. Telegram or WhatsApp is fastest.
- Telegram: @nguyenlnp
- WhatsApp / Zalo: +84 914 859 234
- Email: [email protected]