Sats Coop catalog

Self-hosted LLM (Ollama) hardening

Ollama has no authentication of its own. Keep it off the public internet.

Free excerpt

Check the bind address

ss -tlnp | grep 11434 should show 127.0.0.1. OLLAMA_HOST=0.0.0.0 or -p 11434:11434 in Docker exposes it to everyone who can reach the port.

Test from outside

From another host: curl -s http://YOUR_IP:11434/api/tags. If it answers, anyone can use your GPU and models.

Put a gateway in front

nginx with TLS, API keys, rate limits and an allow-list of endpoints (for example only /api/chat).

Sandbox the service

Run under a dedicated user with systemd hardening options and keep models on a separate volume.

The paid runbook has the gateway config, systemd sandbox, Docker variant, GPU sizing, monitoring and an ollama-audit.sh exposure check.

Get the full pack

Sats Werkstatt