Self-host a private AI assistant on a VPS: pick a server, install Ollama, deploy an agent framework, wire up automation — full step-by-step.

💡 What You Will Learn

Self-host a private AI assistant on a VPS: pick a server, install Ollama, deploy an agent framework, wire up automation — full step-by-step.

📜 Table of Contents

• Why Self-Host
• The Three Components
• Step 1: Pick a VPS
• Step 2: Install Ollama
• Step 3: Deploy an Agent Framework
• Step 4: Add Automation (n8n)
• Cost Comparison
• FAQ

Self-Hosted AI Agent Guide 2026: Run Your Own Private AI Assistant on a VPS

ChatGPT, Claude and Gemini are convenient, but privacy, per-token cost and service changes are real concerns. This guide builds your own assistant on a VPS: open-source models, your data on your machine, no per-token billing.

Why Self-Host

1. Privacy: commercial APIs send your conversations to their servers; sensitive data should not go through them.
2. Cost: API bills grow linearly with usage; self-hosting is a fixed hardware cost.
3. Dependency: price changes, rate limits and outages can break your workflows.

Open-source tools make this easy: OpenClaw (380k+ stars), Ollama, n8n.

The Three Components

Step 1: Pick a VPS

• 7B quantized (daily Q&A): 4-core CPU + 16GB RAM minimum, GPU optional
• Smooth 7B: GPU instance with 8GB+ VRAM
• 70B class: multi-GPU, expensive for personal use Recommendation: a 4-core/16GB CPU instance for 7B quantized, or practice on an old desktop/NAS first. 50GB+ disk (models are 4-10GB each).

Step 2: Install Ollama

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5:7b
ollama run qwen2.5:7b "Hello, introduce yourself"

Expose the API: OLLAMA_HOST=0.0.0.0:11434 ollama serve (restrict source IPs in the firewall).

Step 3: Deploy an Agent Framework

With OpenClaw (see official docs for install): configure the model endpoint to http://localhost:11434/v1, model qwen2.5:7b, then enable tools (web search, code execution, browser) as needed. Your conversations stay on your server.

Step 4: Add Automation (n8n)

n8n is an open-source visual workflow tool: email → agent summary → database; scheduled job → fetch data → agent analysis → push to your IM. Example flow: daily 9:00 → read yesterday's sales → agent writes summary → push to IM.

Cost Comparison

FAQ

Q: Is self-hosted quality comparable to GPT/Claude? A: 7B models handle daily Q&A, summarization, classification; complex reasoning lags flagship commercial models. Go 70B or hybrid (sensitive tasks local, heavy tasks API) for higher ceilings. Q: Is CPU inference slow? A: 7B quantized is ~2-5 token/s — fine for short Q&A, slow for long outputs. GPU for smooth experience. Q: Security notes? A: Ollama binds to localhost by default. If exposed, restrict source IPs, add an API gateway or use Tailscale; never leave port 11434 open to the internet. Q: Updating models? A: ollama pull fetches new versions; update agent/n8n via their upgrade commands after backing up configs.

Note: VPS specs, model versions and install steps change; check official docs.