Skip to content
TechTipsEasy Find a guide

Local AI

How to run local AI with Ollama and check its privacy settings

Run a local model with Ollama, confirm when prompts stay on your machine, inspect active models and avoid unsafe network exposure while testing safely.

By TechTipsEasy · · 4 min read

What local means in Ollama

Ollama is useful when you want a quick local model experiment without building an inference stack. The Ollama quickstart shows the basic flow: install Ollama, open a terminal, choose a model and run it. Local experiments are attractive because they let you test prompts against private drafts, code snippets or notes before deciding whether a hosted model is necessary. Local does not mean every setting is automatically private, though. You still need to check whether you are using a local model or a cloud feature, where the server listens, and what data you paste into prompts.

Run the first local check

After installing Ollama, run a small prompt against a local model you have intentionally downloaded. A safe pattern is:

ollama list
ollama run <model-name>
> Summarize this sentence in five words: local setup check only.

Then open a second terminal and inspect active models:

ollama ps

The Ollama FAQ explains that ollama ps can show processor usage, which helps you see whether a model is using GPU, CPU or both. This is a practical sanity check, not a benchmark. A laptop can feel fast on short prompts and slow on long context. Measure your own task rather than assuming a model name predicts performance.

Check privacy settings

Ollama's FAQ distinguishes local runs from cloud processing. For a local-only setup, set disable_ollama_cloud to true in ~/.ollama/server.json, preserving any existing configuration keys. Restart Ollama and check for its cloud-disabled log message:

{
  "disable_ollama_cloud": true
}

The FAQ also documents OLLAMA_NO_CLOUD=1. Set that variable for the process or service that actually launches Ollama, then restart it; an export in an unrelated terminal will not reconfigure an already running app. Turning off cloud features also disables Ollama's cloud models and web search.

Keep the listener local

Ollama binds to a localhost address by default according to the FAQ. Keep it that way unless you have a reviewed reason to do otherwise. Do not casually expose a local model server to a network just because another device needs access. If a teammate needs the same experiment, give them the reproducible prompt, model name and hardware notes instead of opening a broad listener. Local AI is safest when it stays local.

Example privacy checklist

  • Model source: record the model name you ran and whether it was already downloaded.
  • Prompt data: start with fake customer names, fake secrets and short excerpts.
  • Network binding: keep the server on localhost during experiments.
  • Cloud setting: disable cloud features if the test requires local-only behavior.
  • Output handling: do not paste generated output into production docs without review.

Troubleshooting

If the command fails because no model is found, download the model intentionally rather than trying random names until something works. If responses are slow, reduce prompt length and context before judging the setup. If memory pressure appears, close other heavy applications and try a smaller model. If output quality is weak, create a test set with five representative prompts. One prompt is a vibe check; five prompts with expected traits are the beginning of an evaluation.

When local is enough

Ollama is a strong fit for learning model behavior, drafting offline notes, testing prompt formats, classifying small batches and comparing privacy-sensitive examples before using an API. It is a weaker fit when many users need concurrent access, audit logs, managed scaling or strict uptime. When the experiment becomes a shared service, compare a serving stack such as vLLM or a hosted provider. The best next step is a tiny lab notebook: model, prompt, input length, output quality, latency impression, privacy setting and whether the result was good enough for the task.

Record your environment

Local results depend on hardware, operating system, memory pressure and model size. Write down the machine, model, prompt length and whether other heavy apps were running. These notes prevent a common misunderstanding: one person's fast local demo may be another person's slow laptop experience.

Verify the local API with synthetic input

Once a downloaded model runs, check the local service separately from the chat interface. This request lists the server's model metadata without sending a prompt:

curl http://127.0.0.1:11434/api/tags

Record the model tag you intend to use and test a short synthetic prompt before importing a document. Model downloads and software updates still require a network connection when you obtain them; local inference describes where that run processes the prompt. Also check any separate chat frontend or plugin you add, because its own settings may enable remote services. Keep the listener on localhost throughout this experiment.

Sources and further reading

Documentation checked October 4, 2026. Product features and availability can change.

Keep learning