Twenty minutes, no API key, no account. Install Ollama, pull one model, point GhostNote at it. From that point nothing you ask ever leaves your machine.
Steps and model names last checked on 29 July 2026.
Ollama is a free app that runs open-weight language models locally and exposes them on your Mac at localhost:11434. Get it from ollama.com or, if you use Homebrew, brew install ollama. Either way it installs a small background service — there is no separate step to “start the server.”
Open Terminal and run something like:
ollama pull llama3.2
That downloads the model once, over the internet — the one moment in this whole setup that actually needs a network connection. After it finishes, the model lives on disk and every run after that is fully local. Llama 3.2 is a reasonable first pull on a MacBook Air; Llama 3.3, Mistral, Qwen 2.5 Coder and DeepSeek Coder V2 are worth trying once you know your machine can carry the bigger versions.
In GhostNote, open the provider picker and choose Local (Ollama). There is no key field to fill in — GhostNote just talks to Ollama on localhost:11434. If Ollama is running and a model is pulled, it works immediately, on every plan including the free tier.
This is the part people skip and then wonder why everything is slow. Mac unified memory is shared between the model, the OS, and every other app you have open. As a rough rule of thumb: a small (7–8B parameter) model wants a machine with real headroom above 8GB, and it's competing with Chrome, Slack and everything else in your dock for that space. Bigger models want considerably more. If you're on an 8GB Mac, close the memory-hungry apps before you pull anything larger than the smallest models — and don't expect to run a 30B-class model comfortably at all.
Offline mode isn't a paid feature. It runs on the free tier — 15 minutes a month, every feature working, no card — the same as every other provider option.
Be honest with yourself about the tradeoff. A local 7–8B model is not going to reason through a dense multi-step problem the way a frontier model does — it's genuinely good at quick lookups, definitions, rephrasing, and straightforward code, and genuinely worse at the kind of question where you need the best possible answer, not just an answer. If the stakes are high — a real interview, a technical design review, a call you can't afford to fumble — a cloud model is still the better tool, and GhostNote supports those too with your own Anthropic, OpenAI or Google key.
And if your reason for wanting this is privacy rather than cost, it's worth knowing the invisibility itself doesn't depend on which provider you pick. GhostNote's window is hidden from screen sharing on every plan regardless of whether you're running Ollama or a cloud key — the offline route just removes the network request too.
Cost and control. Ollama is free forever once it's installed, so there's no marginal price for experimenting. And unlike some tools in this space, nothing about running local requires you to accept a EULA that reserves the right to transmit or sell your usage data — a term we noticed in Ofradr's own EULA when we checked it on 29 July 2026. With Ollama there's no data-handling clause to read, because there's no data leaving your Mac to handle.
GhostNote supports Ollama alongside Anthropic, OpenAI and Google — 47 models total, all named, switchable from one menu. See the full breakdown on the offline AI page or the full comparison.