The browser shows a local address. Next to the model, however, a web search is switched on, and an additional service prepares the documents. The first visible screen therefore says little about which stations a request actually passes through. When getting started with local AI, a small map of your own working route is worth the effort.
Ollama’s FAQ distinguishes between local and hosted models. It documents a local-only mode in which Ollama’s cloud features are disabled, as well as the default binding to loopback. These settings are concrete building blocks. A complete statement about additional interfaces and tools does not follow from them.
Three stations instead of a blanket “local” label
Separate the interface, model processing and additional functions. The interface accepts the message. The model generates the answer. Search, document preparation or tools can add further connections. Each station needs its own assignment: on this device, on another computer of your own or with an external service.
A simple text summary is suitable for the first test because it requires neither document search nor web research. A manageable working route is easier to understand than an assistant with numerous extensions active at the same time.
Setting the documented local-only mode
If you have already installed Ollama and prepared a suitable local variant, you can first switch off its own cloud features. The FAQ names the environment variable OLLAMA_NO_CLOUD=1 for this, or an entry in ~/.ollama/server.json:
{
"disable_ollama_cloud": true
}Add the entry so that it fits your existing configuration, instead of replacing an existing file without checking it. Restart Ollama after the change. According to the instructions, the log then shows “Ollama cloud disabled: true”. The platform-specific setup of environment variables is also described in the FAQ.
This setting concerns Ollama’s cloud models and web search. A different chat interface may still have external functions of its own. Disable or examine such extensions separately before you conclude that the whole route is purely local.
A test of your own with known details
Create a short course sheet as plain text: ten participants, a morning session, sealed material only after approval. The task is: “Summarise these details in three factual sentences. Keep the number, the time window and the restriction. Do not add an offer or any unknown responsibility.”
Use an existing local model variant that you have explicitly chosen. Note its exact identifier, the runtime and the interface used. Compare the answer with the source text. Fluent wording is not a suitable summary if the word “only” or the approval condition gets lost.
Then repeat the same task in a new conversation. After that, change exactly one detail, for example the number of participants. This shows you whether the answer uses the current input or carries over a previous version. Only runs that have actually been carried out deliver results; no model quality is predicted here.
What an offline test also clarifies
A controlled test without an internet connection can show whether the small prepared working route still works. For this, the required files must already be present. A missing initial download is a different problem from a function that needs a remote service for every processing step.
The behaviour of one task is also no substitute for a full network analysis. A successful local chat does not automatically prove the data paths of every extension. Anyone who wants to examine these more closely has to look separately at the processes and connections actually in use.
Getting started nevertheless remains pleasantly manageable: a prepared local model, a known text sample and a short list of the stations. This gives you control over the decisive questions without deriving a comprehensive data protection promise from the name of an application.
Later, you can add document search or tools one at a time. After each extension, you update your map. In this way, local AI becomes a traceable set-up whose behaviour you can understand step by step.
Sources and status
Sources as of 25 October 2026. Ollama FAQ — local-only mode.
