How to run a private ChatGPT at home on your own computer (Ollama + Open WebUI)

Short answer: Install Ollama (free) on the computer that will be the server, download one open model that fits its memory (for example ollama pull gemma4:e4b, 6.6-9.5 GB, or gpt-oss:20b, 14 GB, if you have 16 GB+ of RAM or VRAM), then run Open WebUI (free) in Docker to get a ChatGPT-style chat page at http://localhost:3000. Create the admin account straight away, turn on sign-ups for family members, and open http://SERVER-IP:3000 from phones and laptops on the same Wi-Fi; use Tailscale's free Personal plan to reach it away from home without opening router ports. Chats and documents stay on your machine when you use local models. Expect smaller models to be slower and less capable than ChatGPT or Claude. People below served models from a PS5 to every device in the house, ran a model bigger than their laptop's memory and tested speeds before buying a GPU.

Real setups where people served open models from a PS5 to every device in the house, ran a 48 GB model on a 32 GB laptop, tested speeds before buying hardware and kept an always-on agent on a $145 mini PC. 24 matching use cases below, each with a public source.

Looking for something else? Search all 1965 use cases or Get the daily 5.

• What you need
• One computer
• Family Wi-Fi
• Away from home
• Copy-paste prompts
• Safety
• Costs
• Real examples

What you need

Prices are from each official page, checked Oct 26, 2026; they exclude tax and can change. Swipe the table sideways on a phone.

Set it up on one computer

Mac, Windows or Linux. About 20 minutes plus the model download.

1. Install Ollama. Use the official installer, or download the app from ollama.com/download. On Mac and Windows open the Ollama app once; on Linux the installer starts it as a service.
   shellCopy

   # Mac or Linux
   curl -fsSL https://ollama.com/install.sh | sh
   # Windows (PowerShell)
   irm https://ollama.com/install.ps1 | iex
   
   ollama --version

2. Download a model that fits your memory. Pick by RAM (or GPU VRAM): 8 GB, gemma4:e2b; 16 GB, gemma4:e4b; 16 GB+ with a good GPU or a 24-32 GB Mac, gpt-oss:20b. Sizes are from Ollama's library.
   shellCopy

   ollama pull gemma4:e4b        # 6.6-9.5 GB, good start on 16 GB
   # ollama pull gemma4:e2b      # 4.6-7.5 GB, for 8 GB machines
   # ollama pull gpt-oss:20b     # 14 GB, needs 16 GB+ RAM or VRAM
   ollama run gemma4:e4b "Say hello in one sentence."   # type /bye to leave a chat
   ollama ps                      # PROCESSOR column: 100% GPU, CPU or a split

   
   If ollama ps shows mostly CPU and answers crawl, use a smaller model.
3. Start Open WebUI. Install Docker, then run Open WebUI. It finds Ollama on the same computer by itself (Open WebUI quick start).
   shellCopy

   openssl rand -hex 32          # copy the output into WEBUI_SECRET_KEY below
   docker run -d -p 3000:8080 \
     --add-host=host.docker.internal:host-gateway \
     -v open-webui:/app/backend/data \
     -e WEBUI_SECRET_KEY=paste-the-key-here \
     --name open-webui --restart always \
     ghcr.io/open-webui/open-webui:main

4. Or start it without Docker. Open WebUI also installs with pip.
   shellCopy

   # No Docker? Python 3.11 or 3.12:
   pip install open-webui
   open-webui serve              # then open http://localhost:8080

5. Create the admin account. Open http://localhost:3000 right away. The first account created becomes the admin, so do this before anyone else can reach the page. Pick your model at the top of a new chat and say hello.
   shellCopy

   # Linux only, if Open WebUI shows no models:
   sudo systemctl edit ollama.service
   #   add these two lines, save and close:
   #   [Service]
   #   Environment="OLLAMA_HOST=0.0.0.0:11434"
   sudo systemctl daemon-reload && sudo systemctl restart ollama

   
   Linux note: Docker reaches Ollama through host.docker.internal, so Ollama must listen beyond 127.0.0.1. Keep that machine on a trusted home network.

Share it with your household

Every phone, tablet and laptop on your Wi-Fi uses the same server, each with its own login.

1. Find the server's IP address. Then give it a fixed IP (DHCP reservation) in your router so it doesn't change.
   shellCopy

   ipconfig getifaddr en0        # Mac on Wi-Fi
   ipconfig                      # Windows (IPv4 Address)
   hostname -I                   # Linux

2. Let family members sign up. As admin, open your avatar > Settings > Admin > Authentication and turn on new sign-ups. New users wait as Pending until you approve them in the Admin Panel. Turn sign-ups off again once everyone has an account.
3. Open it on phones. On each device go to http://SERVER-IP:3000 (for example http://192.168.1.20:3000), log in, then use the browser's Share > Add to Home Screen so it opens like an app.
4. Make a family assistant preset. Create a model with your own instructions and a larger context window, then pick 'family' in Open WebUI. You can also do this in Open WebUI's Workspace > Models.
   textCopy

   # Save as Modelfile, then: ollama create family -f Modelfile
   FROM gemma4:e4b
   PARAMETER num_ctx 8192
   SYSTEM """You are our family helper. Answer in plain language. Say when you are not sure.
   For medical, legal or money questions, give general information and suggest asking a professional."""

5. Keep the server awake. A sleeping computer means 'server not found' on every phone.
   shellCopy

   caffeinate -s                                  # Mac: stay awake while plugged in (leave Terminal open)
   powercfg /change standby-timeout-ac 0          # Windows: never sleep on AC power
   sudo systemctl mask sleep.target suspend.target  # Linux: disable sleep

Use it away from home with Tailscale

Reach the server from your phone on mobile data without opening any ports on your router.

1. Install Tailscale everywhere. Install Tailscale on the server and on each phone or laptop and sign in with the same account (the free Personal plan allows 6 users). Download from tailscale.com/download.
2. Serve Open WebUI on your tailnet. This gives a private HTTPS address that only your devices can open (Tailscale Serve docs). Don't use Funnel, which publishes to the whole internet.
   shellCopy

   # Install Tailscale on the server and on each phone/laptop, sign in to the same account, then on the server:
   tailscale serve --bg 3000
   # It prints an https://<machine>.<tailnet>.ts.net address that works only on your tailnet
   tailscale serve status
   tailscale serve reset          # stop sharing

3. Keep it updated. Update Open WebUI and models now and then. Your accounts and chats are kept in the open-webui Docker volume.
   shellCopy

   docker rm -f open-webui
   docker pull ghcr.io/open-webui/open-webui:main
   # then run the same docker run command again (your chats live in the open-webui volume)
   ollama pull gemma4:e4b        # re-pull to update a model

I want to run a private ChatGPT-style assistant at home with Ollama and Open WebUI.
My server: [Mac/PC model], [RAM] GB RAM, GPU: [GPU and VRAM, or none], free disk: [GB].
We'll use it for: [homework help / writing / summarizing documents / coding].
Recommend 2-3 models from ollama.com/library that fit with room to spare, with the exact `ollama pull` names, download sizes and what speed to expect. Tell me which to try first and when I'd need a bigger machine.

Set up a private ChatGPT on this computer: install Ollama with the official installer from ollama.com, pull [model], then run Open WebUI in Docker on port 3000 with a volume named open-webui, --restart always and a random WEBUI_SECRET_KEY from `openssl rand -hex 32`.
Check my OS, RAM and whether Docker is installed first. Show me each command before running it and ask before installing anything. Don't open any router ports or enable public sharing. Finish by confirming http://localhost:3000 loads and `ollama ps` shows the model.

Open WebUI at http://localhost:3000 shows no models. Ollama is installed on this [OS] machine and `ollama list` shows [models].
Diagnose step by step: check the Ollama service is running, test http://localhost:11434 from the host and http://host.docker.internal:11434 from inside the open-webui container, check OLLAMA_HOST and the Open WebUI connection settings. Explain the cause, fix it with the smallest change, and keep Ollama off the public internet.

My Open WebUI server is at [server IP]:3000 on my home Wi-Fi. I'm the admin. I want [number] family members, including [kids' ages], to use it on their phones.
Walk me through: turning sign-ups on and approving users, turning sign-ups off afterwards, which models each user can see, a sensible system prompt for kids, and how to check nothing is exposed outside my network. Use plain language.

Audit my local AI setup (Ollama + Open WebUI on [OS]) for anything that sends data off this computer. Check Ollama cloud models (OLLAMA_NO_CLOUD or disable_ollama_cloud in ~/.ollama/server.json), Open WebUI web search, image generation and any external API connections.
List each item with: on or off, what data it would send and where, and the command or setting to turn it off. Don't change anything until I approve.

In Open WebUI, help me set up a Knowledge collection called 'House' with these files: [list: appliance manuals, warranty PDFs, school calendar]. Then create a model preset 'House helper' based on [model] that answers only from that collection, quotes the file and page it used, and says 'not in our documents' when it can't find the answer. Give me 5 test questions to check it works.

Benchmark my Ollama server: run [model] with a 300-word prompt three times and report time to first token and tokens per second from `ollama run --verbose`, plus what `ollama ps` says about GPU vs CPU.
Then tell me whether a smaller model or more RAM/VRAM would help most, and the cheapest change that would make answers feel at least twice as fast. Don't buy or install anything.

Safety checklist

• Never forward ports 3000 or 11434 on your router. The Ollama API has no password; anyone who can reach it can use your models. Use Tailscale Serve (private) instead of Funnel (public).
• Create the Open WebUI admin account immediately after the first start: the first person to sign up becomes admin.
• Keep sign-ups off except when adding someone, and approve each Pending user yourself.
• Set your own WEBUI_SECRET_KEY and keep it private; don't share screenshots of the docker command with the key in it.
• Local models are confidently wrong more often than big cloud models. Check medical, legal, money and homework answers against a trusted source.
• Ollama cloud models and Open WebUI web search send data off your computer. Turn them off (OLLAMA_NO_CLOUD=1) if privacy is the point.
• If you let a coding agent install it, keep its command approval on and read each command, especially anything with sudo.

Related guides: How to run local LLMs with AI agents (Claude Code, Codex and OpenCode on Ollama) · How to self-host AI agents on a Raspberry Pi, Mac mini or homelab · How to compare AI models on the same prompt · How to use AI agents for Linux, servers and homelabs

What it costs

• Ollama: Free. Open source; optional paid cloud models you can switch off
• Open WebUI: Free. Open source; runs in Docker or with pip
• Docker Desktop: Free for personal use. Paid only for larger companies; Linux can use Docker Engine for free
• Tailscale: Free Personal plan. Up to 6 users, unlimited devices, non-commercial use
• Hardware: $0 if you reuse a computer. 16 GB+ RAM or a GPU with 8 GB+ VRAM makes answers much faster; electricity depends on the machine

Real examples (24)

Common questions

How do you run a private ChatGPT at home on your own computer (Ollama + Open WebUI)?

Install Ollama (free) on the computer that will be the server, download one open model that fits its memory (for example ollama pull gemma4:e4b, 6.6-9.5 GB, or gpt-oss:20b, 14 GB, if you have 16 GB+ of RAM or VRAM), then run Open WebUI (free) in Docker to get a ChatGPT-style chat page at http://localhost:3000. Create the admin account straight away, turn on sign-ups for family members, and open http://SERVER-IP:3000 from phones and laptops on the same Wi-Fi; use Tailscale's free Personal plan to reach it away from home without opening router ports. Chats and documents stay on your machine when you use local models. Expect smaller models to be slower and less capable than ChatGPT or Claude. People below served models from a PS5 to every device in the house, ran a model bigger than their laptop's memory and tested speeds before buying a GPU.

Is it really private?

With local models, prompts and answers are processed on your computer and Ollama's docs say they aren't sent to ollama.com. Ollama also offers cloud models, which do leave your machine; set OLLAMA_NO_CLOUD=1 to turn them off. Open WebUI features such as web search send the search query to the provider you configure.

What computer do I need?

Any recent Mac, Windows PC or Linux machine with 8 GB of RAM can run a small model such as gemma4:e2b (4.6-7.5 GB). Apple Silicon Macs with 16 GB+ or a PC with an 8 GB+ NVIDIA or AMD GPU feel much faster. A tool listed below (Llmspeed) lets you feel the speed of a model on a given GPU or Mac before you buy one.

Is it as good as ChatGPT?

No for hard reasoning, long research or current events: small local models know less and make more mistakes. It's good for drafting, summarizing, homework help, brainstorming and private documents. You can keep using ChatGPT or Claude for hard tasks.

Can my family use it on their phones?

Yes. Open WebUI works in a phone browser at http://SERVER-IP:3000 and can be added to the home screen like an app. Each person gets their own login and chat history; the admin approves new accounts.

Does it work without internet?

Yes, once the model is downloaded. Chat works fully offline on your home network; only downloads, updates and optional web search need the internet.

Do I need a GPU?

No. Ollama runs on the CPU too, just slower. `ollama ps` shows whether a model is on the GPU, the CPU or split between them.

Ollama or LM Studio?

Both are free. LM Studio is a desktop app with a model browser and a built-in server (Developer tab or `lms server start`); Ollama is lighter and pairs directly with Open WebUI. Use whichever you find easier; Open WebUI can connect to either through its OpenAI-compatible settings.

Can a PS5 or an old laptop be the server?

An old laptop works with small models. One project below (PS5LM) runs llama.cpp on a jailbroken PS5's GPU and serves models to the whole house; jailbreaking has its own risks and isn't needed for this guide.

Which AI tools do people use to run a private ChatGPT at home on your own computer (Ollama + Open WebUI)?

Across the 24 sourced use cases in this guide, the most used tools are Ollama, Claude Code, OpenAI Codex, Qwen, Claude.

How much does it cost to get started?

2 of the 24 use cases can start on a free plan. The most common paid starting price is $17/mo (8 use cases). Prices come from each tool's official pricing page.

How long does setup take?

10 of the 24 use cases list a setup time of an hour or less. Each card shows the setup time the builder reported.

Are these real examples?

Yes. Every use case links to its public source (an X post, a Hacker News thread or a GitHub repo) with the author and date, so you can check it yourself.

More how-to guides

How to make faceless YouTube videos with AI · How to automate work with n8n and AI · How to vibe code an app with Claude Code · How to build AI agents for SEO and ads · How to run multiple coding agents together · How to build an MCP server for AI agents · How to add memory to coding agents · How to secure vibe-coded apps and agents · How to observe and debug AI agents · How to build agent memory and context layers · How to control coding agents from your phone · How to add payments to vibe-coded apps · How to run local LLMs with AI agents (Claude Code, Codex and OpenCode on Ollama) · How to set up Grok Bot agents for real work · How to build with Gemini 4 Argon and Gemini agents · How to use AI agents for taxes, bills and budgets · How to make AI video ads and Reels with agents · How to set up Hermes Agent workflows · How to vibe code a game with AI · How to replace SaaS subscriptions with vibe-coded tools · How to use AI agents for errands, calls and shopping · How to use Cursor agents and automations · How to make 3D scenes and models with AI and Blender · How to use Claude Code skills and plugins · How to cut AI coding agent costs and token use · How to automate a small business with AI agents · How to compare AI models on the same prompt · How to build hardware projects with AI agents · How to make AI music videos and songs · How to use AI agents with Obsidian and your notes · How to build apps and games for your kids with AI · How to use ChatGPT for real projects and everyday tasks · How to make AI agents ask for approval before they act · How to build AI trading and investing agents · How to build AI agents that make and answer phone calls · How to plan meals and groceries with AI agents · How to use AI agents for fitness and health · How to undo, checkpoint and stop runaway coding agents · How to use AI agents for job hunting · How to automate social media posts with AI agents · How to build a smart home and voice assistant with AI agents · How to make money with AI agents · How to build a website with AI agents · How to use AI agents for research and science · How to build an iPhone or Android app with AI · How to automate email and your inbox with AI agents · How to use AI agents for Linux, servers and homelabs · How to build Telegram, Discord and WhatsApp bots with AI · How to research memecoins and crypto with AI agents · How to remove ads and bloatware from your smart TV with AI · How to self-host AI agents on a Raspberry Pi, Mac mini or homelab · How to build and install your own app on a Samsung TV with AI · How to fix a printer that won't print with AI · How to learn anything with an AI tutor · How to build an AI second brain and personal wiki · How to reverse-engineer and revive old devices with AI · How to edit videos with Claude Code and Codex · How to set up an AI family assistant · How to make a Godot, Unity or 3D game with AI agents · How to build a custom Android TV or Google TV launcher with AI · How to use Claude Code and Codex together · How to build your own invoicing and bookkeeping app with AI · How to make a browser arcade game with AI · How to make a product demo or launch video with AI · How to fix Wi-Fi and secure your home network with AI · How to control Claude Code and Codex by voice · How to design a PCB and electronics with AI · How to save, share and resume coding agent sessions · How to run AI coding agents in a sandbox · How to make a multiplayer game with AI · How to let AI control your computer (computer use) · How to build a Mac menu bar app with AI · How to port code to another language with AI · How to make an AI animated short film or anime · How to build your own coding agent or harness · How to get notified when Claude Code or Codex needs you · How to build a Chrome extension with AI · How to use AI agents with Excel and Google Sheets · How to run Claude Code and Codex in a desktop app (T3 Code, Codex app, Claude Desktop) · How to make a song or album with AI · How to vibe code classroom apps and quizzes as a teacher · How to plan and book a trip with an AI agent · How to make motion graphics and animated videos with Claude Code · How to mod and reverse engineer games with AI · How to run AI agents overnight while you sleep · How to negotiate bills and get refunds with an AI agent · How to let an AI agent use your phone (iPhone and Android) · How to build a Polymarket bot with AI (and Kalshi) · How to use Claude Duet: real examples and tips · How to build private apps just for yourself with AI · How to design and 3D print things with AI agents · How to automate customer support with AI agents: real setups · How to find leads and automate sales outreach with AI agents · How to run a Shopify or online store with AI agents · How to launch an app and get your first users with AI · How to build a team of AI agents with a chief of staff · How to make a hand-tracking or gesture-controlled game with AI · How to test your app with AI agents before you ship · How to give an AI agent a budget or a card safely · How to share passwords and logins with an AI agent safely · How to get AI code review on your pull requests · How to review contracts, leases and fine print with AI · How to make slides and presentations with AI · How to take meeting notes and track follow-ups with AI · How to scrape websites and extract data with AI agents · How to make a podcast with AI · How to use AI agents for security scans and threat intel · How to build a dashboard with AI agents · How to clone your voice and make AI voiceovers · How to make AI images, thumbnails and product photos · How to write a CLAUDE.md or AGENTS.md rules file · How to build an AI chatbot for your own documents · How to port or remake a classic game with AI · How to build a VST plugin or synth with AI · How to trigger AI agents with webhooks and events · How to find a home or rental with AI agents (Zillow in Claude and ChatGPT) · How to track Claude Code and Codex usage limits · How to build a daily AI briefing or news digest · How to check if your devices are spying on you with AI · How to track smart-money wallets and copy trade with AI · How to translate and localize your app or content with AI · How to write and publish a book with AI · How to control your TV with an AI agent · How to take your robot vacuum off the cloud with AI · How to connect an AI agent to Home Assistant · How to let an AI agent do your bookkeeping (Xero and QuickBooks) · How to run an AI agent 24/7 on an old Android phone or TV box · How to see all your bank accounts in one place with AI