A self-hosted AI assistant runs on hardware you own. Your Mac Mini. Your VPS. Your Raspberry Pi. No data leaves your network unless you say so. No monthly subscription to some cloud platform that changes pricing whenever they feel like it.
I run my entire business on a self-hosted AI setup. 13 agents, always on, handling content, research, email, scheduling. All running on a Mac Mini sitting on my desk in Bali. Total cloud cost: $0.
Here's how to set up your own.
What's in this guide
Why Self-Host Your AI Assistant
Three reasons people go self-hosted:
1. Data ownership. When you use ChatGPT or any cloud AI, your prompts and data hit someone else's servers. For personal use, maybe fine. For business data, client information, financial records? That's a different conversation. Self-hosting means your data stays on your machine. Period.
2. Cost control. Cloud AI subscriptions add up. $20/month here, $200/month there. A self-hosted setup has a one-time hardware cost, then you pay only for electricity and whatever API calls you choose to make. My Mac Mini costs roughly $8/month in power.
3. Customization. Cloud platforms give you what they give you. Self-hosted means you pick your models, your integrations, your rules. Want your assistant to manage your Telegram, control your smart home, and write your newsletter? You build exactly that.
Who is this for? Founders, freelancers, and small teams who want an AI assistant that actually does things (not just chat). If you just need to ask questions, ChatGPT is fine. If you want an assistant that runs 24/7, manages your tools, and keeps your data private, keep reading.
Best Hardware for a Self-Hosted AI Assistant
You don't need a $5,000 server. Here's what actually works in 2026:
| Hardware | Cost | Best For | Can Run Local LLMs? |
|---|---|---|---|
| Mac Mini M2/M4 | $499-$799 | Always-on home server, best all-rounder | Yes (7B-13B models comfortably) |
| VPS ($5-20/mo) | $60-240/year | Remote access, no hardware to manage | API-only (not enough RAM for local models) |
| Raspberry Pi 5 | $80-120 | Ultra-budget, low power | Small models only (slow) |
| Old laptop/desktop | $0 (use what you have) | Getting started for free | Depends on specs |
The Mac Mini is the most popular choice in the OpenClaw community. Low power consumption (around 5-7 watts idle), dead silent, runs macOS so you get native Apple integrations. The M2 with 16GB handles most workloads without breaking a sweat.
Budget option: Oracle Cloud offers a free tier with 4 ARM CPUs and 24GB RAM. It's enough to run OpenClaw with cloud API models. No hardware purchase needed. The signup process can be finicky, but once you're in, it's genuinely free forever.
Setting Up OpenClaw as Your Self-Hosted Assistant
OpenClaw is open-source and built specifically for self-hosting. It installs via npm, runs on Node.js, and connects to your messaging apps (Telegram, Discord, WhatsApp) so you can talk to your assistant from your phone.
Here's the quick version:
Step 1: Install Node.js on your machine (v20 or higher).
Step 2: Run npm install -g openclaw
Step 3: Run openclaw init to set up your config.
Step 4: Add your API key (Anthropic, OpenAI, or local Ollama).
Step 5: Connect a messaging channel (Telegram bot is the fastest).
Step 6: Start it: openclaw gateway start
That's it. You now have a self-hosted AI assistant running on your own hardware, accessible from your phone.
For a full walkthrough, check installopenclawnow.com for step-by-step install instructions.
Running Local AI Models with Ollama
Want to go fully private? No API calls, no data leaving your machine at all? You need local models.
Ollama is the easiest way to run open-source LLMs locally. It supports models like Llama 3, Mistral, Gemma, and dozens more. Install it, pull a model, point OpenClaw at it.
Hardware reality check:
- 8GB RAM: Can run small models (3B parameters). Good for simple tasks.
- 16GB RAM: Runs 7B models comfortably. This is the sweet spot for most people.
- 32GB+ RAM: Runs 13B-30B models. Noticeably smarter responses.
- 64GB+ RAM: Can handle 70B models. Approaching cloud-quality output.
Apple Silicon Macs are particularly good at this because of unified memory. A Mac Mini M2 with 16GB can run Llama 3.1 8B at around 30 tokens per second. Fast enough for real-time conversation.
Honest take on local models: They're good for privacy and simple tasks. But for complex reasoning, long-form writing, or coding, cloud models like Claude or GPT-4 are still significantly better. Most self-hosters use a hybrid approach: local models for quick tasks, cloud APIs for heavy lifting.
Cloud APIs vs. Fully Local: The Real Trade-Offs
This is the decision every self-hoster faces. Here's the honest breakdown:
| Factor | Cloud APIs (Anthropic, OpenAI) | Fully Local (Ollama) |
|---|---|---|
| Intelligence | Best available models | Good but not top-tier |
| Privacy | Data goes to provider servers | 100% on your machine |
| Speed | Fast (dedicated GPU clusters) | Depends on your hardware |
| Cost | Pay per token ($5-15/month typical) | Free after hardware cost |
| Reliability | 99.9% uptime | Depends on your setup |
| Setup | Add API key, done | Install Ollama, download models |
My recommendation: start with cloud APIs. Get your assistant working, build your workflows, then add local models for specific tasks where privacy matters most. You don't have to choose one or the other. OpenClaw lets you use both simultaneously.
Security and Privacy Best Practices
Self-hosting gives you control, but also responsibility. A few things to get right:
Keep your machine updated. OS patches, Node.js updates, OpenClaw updates. This is basic but people skip it.
Use SSH tunneling or Tailscale for remote access. Don't expose your assistant's port directly to the internet. Tailscale creates a private network between your devices. Free for personal use, takes 5 minutes to set up.
API keys are secrets. Store them in your OpenClaw config (which lives on your machine). Never commit them to Git. Never paste them in public chats.
Firewall basics. If you're on a VPS, only open the ports you need (SSH on 22, and that's usually it if you're tunneling). On a Mac Mini at home, your router's NAT handles this by default.
Quick security win: OpenClaw has a built-in healthcheck skill that audits your machine's security posture. Run it after setup to catch any obvious gaps.
Real Cost Breakdown: Self-Hosted vs. Cloud
Let's do the math for a year of running an AI assistant:
| Setup | Year 1 Cost | Year 2+ Cost |
|---|---|---|
| ChatGPT Plus | $240/year | $240/year |
| Claude Pro | $240/year | $240/year |
| OpenClaw on Mac Mini (cloud APIs) | $599 hardware + ~$120 API + ~$96 power = ~$815 | ~$216/year |
| OpenClaw on VPS (cloud APIs) | ~$60 VPS + ~$120 API = ~$180 | ~$180/year |
| OpenClaw fully local (Ollama) | $599 hardware + ~$96 power = ~$695 | ~$96/year |
The self-hosted route costs more upfront (if you buy hardware) but gets cheaper every year. More importantly, you get an assistant that actually does things: manages your email, posts to social media, monitors your business, runs 24/7. ChatGPT and Claude Pro are chatbots. A self-hosted OpenClaw setup is an employee.
That difference matters when you're running a business.
OpenClaw Lab is the #1 community for founders building AI agent systems. I share the exact playbooks, skill files, and workflows inside. Weekly lives, expert AMAs, and 265+ members building real systems.
Join OpenClaw Lab →