Self-host AI apps on a VPS: Open WebUI, OpenClaw, n8n
How to self-host AI apps on a VPS: a private ChatGPT with Open WebUI, OpenClaw on Telegram or Zalo, AI workflows with n8n, and how to run them safely.

Key takeaways
- Self-hosting AI apps keeps chats, documents and workflows on your own server, gives the whole team one place to work and replaces per-seat subscriptions with a fixed server cost plus API usage.
- Open WebUI is a private ChatGPT-style chat for a team, OpenClaw is an assistant you reach from Telegram, Zalo or WhatsApp, and n8n runs automations with AI steps.
- Open WebUI supports both OpenAI-compatible APIs and local models through Ollama, but local models need far more RAM and ideally a GPU, so most small VPS setups use API keys.
- Flowise ended in August 2026 and gets no more security fixes, so new projects should start on a maintained tool.
To self-host AI apps on a VPS, run each app in its own Docker container behind nginx with HTTPS on your own domain, connect it to a model provider with your API key, and lock down sign-ups before you share the link. The three most useful apps for a small team are Open WebUI (a private ChatGPT-style chat), OpenClaw (an AI assistant you talk to from Telegram, Zalo or WhatsApp) and n8n (workflow automation with AI steps). Your conversations and data stay on your server, everyone shares one installation, and you pay a fixed price for the server plus what you actually use at the model provider.
This hub explains why to self-host, what each app is for, what resources they need, API keys versus local models, how to keep them isolated and safe, and where to find the detailed guide for each one.
Why self-host AI apps instead of using SaaS?
Hosted AI products are convenient. Self-hosting makes sense when one of these matters to you:
- Your data stays on your server. Chat history, uploaded documents, workflow data and bot sessions are stored in a folder on your VPS, not in another company's workspace. The prompts you send to a model provider still go to that provider, so read its data policy; but what is stored, and for how long, is up to you.
- A flat cost. You pay for the server every month whatever happens. Model usage is billed by your provider per use, which for a team of occasional users is often less than one subscription per person. Set a spending limit at the provider.
- Team access in one place. One Open WebUI with accounts for everyone, one n8n for all automations, one assistant bot for the whole office, with an admin deciding who gets in.
- Your choice of models. Connect OpenAI, Anthropic, DeepSeek, OpenRouter or any OpenAI-compatible provider, and switch without moving your data.
The trade-off: you run the server. Updates, backups and security are your job, which is what the rest of this article is about.
Which AI apps can you self-host on a VPS?
| App | What it is for | Who uses it | Detailed guide |
|---|---|---|---|
| Open WebUI | A private ChatGPT-style chat with accounts, chat history, document upload and per-model settings | A team that wants one shared chat app on its own domain | Self-host ChatGPT with Open WebUI |
| OpenClaw | An AI assistant gateway that connects a model to the chat apps you already use, with a web Control UI | Anyone who wants an always-on assistant on the phone | OpenClaw on Zalo and Telegram |
| n8n | Workflow automation: a trigger starts a chain of steps that call APIs, transform data and send messages, AI steps included | Operations, marketing and anyone replacing Zapier or Make | Self-host n8n |
| Flowise | A visual builder for LLM chatbots and agents; the project ended in August 2026 | Existing installs only | Self-host Flowise and its sunset |
Open WebUI: a private ChatGPT for your team
Open WebUI's documentation calls it "a self-hosted AI platform" with "support for Ollama and OpenAI-compatible APIs". In practice it gives you a familiar chat screen, user accounts with roles, chat history, document upload for questions over your own files, and groups that control who sees which models. It does not contain a model itself; it is the interface in front of one.
The first account created becomes the administrator, so create it right after installing and decide how sign-ups work (for example, new users wait as pending until an admin approves them). The Open WebUI guide covers Docker, nginx and these settings step by step.
OpenClaw: an AI assistant on Telegram, Zalo or WhatsApp
OpenClaw describes itself as "an open-source AI assistant that runs on your own hardware and meets you in every chat app you already use". Its documentation lists channels such as Telegram, WhatsApp, Zalo, Discord, Slack, Signal and Microsoft Teams. You chat with the assistant from your phone, and the conversations, sessions and memory are kept on your server. Access to its Control UI is protected by a gateway token and by approving each new browser.
For Vietnamese teams, the Zalo channel is the interesting one; OpenClaw marks it as experimental. See OpenClaw on Zalo and Telegram.
n8n: automation with AI steps
n8n is a visual workflow tool, a self-hosted alternative to Zapier or Make. Its AI documentation says you can "build AI-powered workflows in n8n with LLM providers, agents, tools, memory, and MCP servers", with providers such as OpenAI, Anthropic and Google. Typical uses: summarise new support emails, classify leads from a form, draft replies, or post a daily report to a chat group. Self-hosted, there is no per-task fee and the workflow data stays with you. See self-host n8n.
What about Flowise?
Flowise was a popular visual builder for LLM apps, but its sunset page announced a code freeze on 29 July 2026, a public archive of the repository on 10 August and the end of life on 31 August 2026. The code stays available under Apache 2.0, but there is no maintainer shipping security fixes. If you already run it, keep it private and plan a move; for new projects, n8n's AI nodes or Open WebUI are the maintained options. The details are in what the Flowise sunset means.
How much RAM and CPU do self-hosted AI apps need?
The apps' own documentation gives little in the way of hardware minimums, so treat these as starting points:
- Open WebUI reports that its
:slimimage used about 300 MiB of RAM at idle in the authors' test, and its quick start warns that this is not a minimum requirement. Document search loads embedding models, which adds memory. - OpenClaw lists Node.js and an API key as requirements in its documentation, without CPU or RAM figures.
- n8n's hosting documentation does not state memory or CPU requirements; workflows that hold many items or large files in memory need more.
For comparison, ZoPanel's App Store, which has tested these apps, gives each one a memory limit and one CPU: 2048 MB for Open WebUI, 2048 MB for OpenClaw and 1024 MB for n8n, and it checks free disk before installing (about 7.9 GB for Open WebUI, 5.8 GB for OpenClaw and 2.6 GB for n8n). Running all three plus the operating system therefore fits a VPS with several gigabytes of RAM to spare. Measure your own use with docker stats before buying a bigger server.
Should you use API keys or local models?
| API keys (OpenAI, Anthropic, DeepSeek, OpenRouter…) | Local models (for example through Ollama) | |
|---|---|---|
| Server needed | A small VPS is enough; the model runs at the provider | Much more RAM, and ideally a GPU |
| Quality and speed | The strongest models, fast | Depends on the model your hardware can hold |
| Cost | Pay per use at the provider | No API bill; you pay for a bigger server |
| Privacy | Prompts go to the provider | Prompts never leave your server |
Open WebUI supports both: its quick start explains connecting OpenAI-compatible providers in Admin Panel → Settings → Connections, and running Ollama alongside it or pointing OLLAMA_BASE_URL at an Ollama server. On a typical VPS without a GPU, small local models run, but slowly, so most teams start with API keys and try local models on a separate, stronger machine.
Whatever you choose, never expose an Ollama port (11434) to the internet without protection, and set a monthly spending limit on every API key, because every user's chats and every automation run are billed to it.
How do you keep self-hosted AI apps secure?
AI apps hold valuable things: API keys that cost money, chat history, documents and bot tokens. The basics:
- One container per app, with limits. A crash or a leak in one app should not reach the others or the server.
- Bind ports to 127.0.0.1 and let nginx serve the app with HTTPS. Ports Docker publishes publicly can bypass a UFW firewall (see Docker apps on a VPS).
- Create the admin account first, then turn off open sign-ups or require approval.
- Keep tokens secret. OpenClaw's gateway token lets someone request access; approve only browsers you recognise.
- Back up the data folders, and store copies encrypted: they contain your API keys.
- Update when security releases come out, and retire apps that no longer get fixes.
- Harden the server itself: SSH keys, a firewall, automatic security updates (secure an Ubuntu VPS).
Self-host AI apps with ZoPanel
ZoPanel's App Store installs Open WebUI, OpenClaw and n8n in one click on the domain you choose, with Let's Encrypt SSL. According to the App Store documentation, every app runs in its own isolated container:
- User namespaces (
userns-remap): user IDs inside the container map to a range no hosting account uses, so root in the container is nobody on the host. - Ports are published on 127.0.0.1 only and reached through the website's nginx; containers cannot talk to each other.
- Each container runs with
no-new-privileges, a process limit and a memory and CPU limit, inside the owning account's resource limits. - App data lives in
/var/lib/zopanel-apps/<instance>/, readable only by root and the container, and each app has a Backups card with Back up now, a schedule and restore. - Containers are cut off from the server's loopback and private networks and from the cloud metadata service.
Two things to know. First, that network isolation means an Ollama or other model server on the same VPS or a private network is unreachable from these apps: ZoPanel turns Open WebUI's Ollama API off, and it does not install Ollama. Use public model APIs, or a model server on another machine with a public address protected by a key. Second, Open WebUI and n8n start behind a setup lock so you create the admin account before anyone else can, and ZoPanel never updates an app by itself: an administrator presses Update to latest, which backs up the data first. Flowise is no longer offered for new installs since ZoPanel 0.4.25.
ZoPanel does not replace a GPU server, and you still run the VPS. The Free plan covers 10 websites, 10 databases and 3 hosting accounts, and each app counts as one website. Guides: Open WebUI, OpenClaw, n8n, pricing. If you also build your own AI apps, see hosting for vibe coding.
Conclusion
Self-hosting AI apps is now practical on a single VPS: Open WebUI for team chat, OpenClaw for an assistant in the chat apps people already use, and n8n for automations with AI steps. API keys keep the server small; local models trade a bigger server for prompts that never leave it. Isolate each app, bind it to localhost behind HTTPS, lock sign-ups, back up its data and keep it updated. Start with the app your team will use first, then follow its detailed guide.
FAQ
Can I run a private ChatGPT on my own server?
Yes. Install Open WebUI on a VPS behind HTTPS and connect it to an OpenAI-compatible API with your key, or to local models through Ollama on hardware strong enough for them. Chats and accounts are stored on your server. The Open WebUI guide has the steps.
Do self-hosted AI apps need a GPU?
Not when they use API keys: the model runs at the provider and the app only needs a normal VPS. A GPU matters when you run models locally, for example with Ollama, because those need much more memory and compute.
Is self-hosting AI cheaper than ChatGPT or Claude subscriptions?
It can be for a team of occasional users, because you pay one server plus per-use API costs instead of a subscription per person. Heavy users may cost more through the API than a subscription. Set spending limits and compare your actual usage.
Is my data private when I self-host an AI app?
What the app stores, such as chats, documents and workflow data, stays on your server. Prompts sent to a model provider through an API still reach that provider, under its data policy. Only local models keep prompts entirely on your machine.
Should I still use Flowise in 2026?
Not for new projects. The Flowise team ended the project on 31 August 2026 and archived the repository, so no security fixes are planned. Keep existing installs private and plan a move to a maintained tool such as n8n or Open WebUI.
Which self-hosted AI app should I start with?
Start with the problem: team chat with AI points to Open WebUI, an assistant reachable from Telegram or Zalo points to OpenClaw, and automating repetitive work across services points to n8n. All three can run on the same VPS.
Put your servers on ZoPanel
Start free with 10 websites on a fresh Ubuntu or Debian server, or talk to our team about moving from cPanel, DirectAdmin or aaPanel and about licences for your business.
Free for 10 websites · No credit card · Installs in minutes
Keep reading

Run your own AI assistant on Telegram, Zalo and WhatsApp with OpenClaw
How to self host OpenClaw on a VPS and talk to your own AI assistant from Telegram, Zalo or WhatsApp: requirements, setup, costs and security.

How to self-host Flowise on a VPS in 2026, and what its sunset means
Self-host Flowise on a VPS with Docker and HTTPS, protect API keys, back up flows, and understand the project's August 2026 end of life before you build on it.

How to self-host n8n on a VPS with Docker, HTTPS and backups
Self-host n8n on a VPS with Docker behind nginx and HTTPS, correct webhook URLs, workflow backups and safe updates, or install it in one click with ZoPanel.