Ollama, a model on your own Mac
Ollama is a free program that runs language models on your own Mac, and the Ollama section of Settings is where Vixlo looks for it, says whether it found it, and lets you choose which installed model the edit page's assistant should use. You do not need Ollama to use Vixlo; without it, the assistant is the one thing that needs another route.
The Local daemon row
The card's title is Local daemon. At its right-hand end a dot and a word say what Vixlo found at the address below:
| Word | Meaning |
|---|---|
| Checking… (grey) | Vixlo is asking. The note reads Looking for a daemon at the address below… |
| Not running (red) | Nothing answered. The note names the address it tried and what to do: Nothing answered at http://localhost:11434 — install Ollama and run “ollama serve”, or fix the address below, then Recheck. |
| Connected (green) | Ollama answered. The note counts what it has: Serving 3 installed models., or Running with no models installed — install one with “ollama pull”. |
Vixlo asks once when it starts, and again only when you ask it to. When Ollama answers with a model Vixlo can run, the edit page's Ask for an adjustment… field sends your sentences to it. When Ollama answers with no models, or with none that can use tools and none that you chose on the Ollama model row, the field is disabled and nothing is sent. The edit page takes this same answer when it opens.
The address
The field under the row holds the address Vixlo asks, and starts as http://localhost:11434, which is where Ollama listens on this Mac unless you have changed it. Edit it if yours is elsewhere, then press ↩ in the field or the Recheck button; either asks again at the new address, and the dot reads Checking… until there is an answer. The address is remembered as you type it, but typing alone asks nothing.
At a localhost address, Ollama is on this Mac and your sentences never leave it. If you point the address at another computer, your sentences go to that computer, and they do so whether or not Local only is on: that switch turns off the cloud providers, not this address.
Recheck is also the button to press after you have started Ollama or pulled a new model: Vixlo does not poll, so until you ask again it is showing the answer it last got.
An edit page that is already open never takes a new answer from Ollama. Starting Ollama or pulling a model shows in this window only when you press Recheck (or ↩ in the address field), and even then only this window shows it: the edit page's field goes on acting on the answer it took when it opened. Press Recheck, then leave the edit page and open it again, and the field has the new answer too.
The Ollama model row
Once Ollama reports at least one installed model, a second row appears under the address, headed Ollama model, with one chip per model and an Automatic chip first.
The order of the chips is Ollama's own answer about each model, not a list inside Vixlo. Models Ollama says can use tools come first; then any it said nothing about; last, those it says cannot, each marked no tool use under its name. The assistant asks a model for one adjustment, and a model can give one only if it can use tools; a model that cannot use tools moves nothing, and Ollama usually refuses the request outright.
- Leave Automatic chosen and Vixlo takes the first model Ollama reports as able to use tools. Only on an older Ollama that reports no capabilities for any model does it fall back to the first installed model, because "did not say" is not "cannot".
- Choose a model and your choice wins over every fallback, even if Ollama would have ranked another first, and even if it is marked no tool use. Such a model is run as you asked, but expect nothing to move: Ollama usually refuses the request, and the edit page shows a card headed The model did not answer with a line like The model call failed: HTTP 400.
The line under the chips states what will actually be used, which is not always what you chose:
- Sends to qwen2.5:14b — Vixlo's choice. under Automatic.
- Sends to qwen2.5:14b. when your choice is installed.
- llama3.2 is not installed right now, so this sends to qwen2.5:14b. Your choice is kept — pull it again and it comes back. when you chose a model that is no longer there. The choice is not forgotten; pulling the model again restores it without your choosing twice.
- No installed model reports tool use, so the edit page's assistant has nothing to run. when nothing installed can do the job and nothing you chose is installed. The assistant's field is disabled for the same reason, and its sentence names models that do work: llama3.1, qwen2.5 and mistral-nemo.
A running Ollama with a model Vixlo can run is used first, whatever cloud keys are saved and whether or not Local only is on. But Vixlo asks Ollama only when it starts and when you ask again: at a later launch where Ollama is not running, a saved Anthropic key with Local only off sends your sentences to Anthropic instead.
If you want your sentences to stay on this Mac, install Ollama on this Mac, pull one of the models above, keep the address at localhost, and turn Local only on. That switch is the one switch that stops a sentence going to a cloud provider whatever keys are saved. Then press Recheck and open the edit page.