DCA Hardened Private LLM Server for Ollama — Support & Quick Start

At a glance

Quick start

This is the listing's usage information, verbatim. Every command in it is run by our QA on a freshly launched instance before an image version can publish, so the text and the tested procedure cannot drift apart.

After launch:

1. Connect over SSH as the opc user with the key you supplied at launch:

   ssh opc@<public-ip>

2. Ollama runs as a systemd service on loopback only (127.0.0.1:11434). It has no authentication of its own, so it is never exposed: nginx fronts it on port 8443 with TLS and an API key, both generated uniquely for this instance at first boot, and the proxy also listens on loopback until step 5. No model ships in the image. Check both services:

   sudo systemctl status ollama nginx

3. Pull a model and run it on the instance. smollm2:135m (271 MB, Apache-2.0) answers in seconds on 2 OCPUs; other models from ollama.com/library run the same way under their own licences, sized to the instance's memory:

   ollama pull smollm2:135m
   ollama run smollm2:135m "Why is the sky blue? Answer in one sentence."

4. Read the API key (root only), then call the OpenAI-compatible API through the proxy, trusting this instance's own certificate authority:

   sudo cat /root/.ollama_default_credentials
   curl --cacert /etc/nginx/ollama-proxy/ca.crt -H 'Authorization: Bearer <api-key>' -H 'Content-Type: application/json' https://localhost:8443/v1/chat/completions -d '{"model":"smollm2:135m","messages":[{"role":"user","content":"Say hello."}]}'

   From your workstation (with a copy of ca.crt), the same https://localhost:8443 works through an SSH tunnel without exposing anything:

   ssh -L 8443:127.0.0.1:8443 opc@<public-ip>

5. To reach the API from your application tier, expose the proxy (never Ollama itself) and restart nginx:

   sudo sed -i 's/listen 127.0.0.1:8443 ssl;/listen 8443 ssl;/' /etc/nginx/conf.d/ollama-proxy.conf && sudo systemctl restart nginx

   Port 8443 is already permitted in the host firewall; 11434 stays blocked. Restrict 8443 in your VCN security list to the application tier's CIDR; never expose it to the internet. Clients use https://<private-ip>:8443/v1 as the OpenAI base URL, or https://<private-ip>:8443 for Ollama's own API (for example /api/embed with an embedding model such as all-minilm), send the key as a Bearer token, and trust /etc/nginx/ollama-proxy/ca.crt (the certificate names localhost, the instance's hostname and its primary private IP). To use your own certificate instead, replace server.crt and server.key in that directory and restart nginx.

6. The API key grants the whole Ollama API, including pulling and deleting models. Rotate it at any time; the old key stops working at once and the new one is written to the credentials file:

   sudo /usr/local/sbin/ollama-rotate-api-key

7. Models live in /var/lib/ollama/models; free space with ollama rm <model>. For anything beyond evaluation, attach a block volume, stop the service (sudo systemctl stop ollama), mount the volume at /var/lib/ollama, run sudo chown -R ollama:ollama /var/lib/ollama && sudo restorecon -R /var/lib/ollama, then start it again. Ollama's cloud features (remote inference and web search) are disabled with OLLAMA_NO_CLOUD=1, so prompts never leave the instance; pulls from the Ollama model registry still work.

The image is CVE-patched at build time. Apply ongoing operating-system and nginx updates with:

sudo dnf -y update. Ollama itself is updated by launching a newer image version; keeping models on a block volume (step 7) carries them across.

What the image provides

• Ollama 0.34.4, the newest stable release, installed from the project's official release archive with its published SHA-256 checksum verified at build time, and running as a dedicated unprivileged service in a locked-down systemd sandbox. The image carries the CPU runtime only: the release's GPU libraries are not installed.
• An authenticated, encrypted endpoint in front of an API that has no authentication of its own. Ollama listens on loopback only and is never exposed; nginx from the Oracle Linux 9 AppStream repository serves both APIs on port 8443 over TLS and admits only requests that carry this instance's API key. The API key and the TLS certificate authority are generated uniquely for each instance on first boot and surfaced to the operator over SSH — never shared, never baked into the image — and the key can be rotated with one command.
• Private by default: Ollama's cloud features (remote inference and web search) are disabled, so prompts never leave the instance. No model weights ship in the image; you pull the models you choose, under their own licences, into a model store that moves to a block volume in one documented step.
• Safe-by-default network posture: the proxy also listens on loopback until you expose it with one documented setting. Port 8443 is declared in the host firewall; the Ollama port never is.
• SELinux enforcing with the proxy confined: every installed file keeps the label the targeted policy expects, and the proxy reaches Ollama through a single port label and the reverse-proxy boolean Red Hat documents, not blanket network access.
• Security hardening aligned to the Oracle Cloud Marketplace image standards: host firewall default-deny with only SSH and 8443 permitted, no account with a usable password, no stray keys, and the image fully patched at build time.

Still stuck?

Email support@dcassociatesgroup.com (response within 1 business day) or use the contact form. Include the listing name, your OCI region, the instance OCID and the output of sudo journalctl -u <service> -n 100 for the unit named in the quick start.

Ollama is a trademark of Ollama, Inc. This image is an independent hardened distribution and is not affiliated with, endorsed by, or sponsored by Ollama, Inc. Models you download are provided by third parties under their own licence terms.