How to run AI offline on Windows
No account, no API key, no subscription, and nothing sent to anyone else's servers. Here's what running AI locally on Windows genuinely gets you — and the parts that still need an internet connection, which most guides quietly skip.
Last updated: September 2026
Why bother running it locally
Cloud assistants are more capable than anything you can run at home. The reasons to run locally anyway are specific:
- Privacy. Screenshots and questions are processed on your machine. Nothing is transmitted, so nothing can be retained, logged or trained on.
- Cost. No monthly fee and no per-token billing. You pay in disk space and electricity.
- It keeps working. On a plane, on hotel wifi, behind a restrictive corporate network, or when the provider has an outage.
- No rate limits. Ask as much as you want, as often as you want.
What actually runs offline — and what doesn't
This is where most guides oversell. An honest breakdown of a screen-aware assistant running locally:
| Capability | Offline? | Runs via |
|---|---|---|
| Answering questions | Yes | Ollama, on your CPU or GPU |
| Reading your screen | Yes | A local vision model |
| Understanding speech | Yes | Whisper, running locally |
| Speaking answers aloud | Partly | Windows' built-in voices work offline |
| Searching the web | No | Needs internet by definition |
Setting it up
Install Ollama
Ollama is the engine that runs models locally. Grab it from ollama.com/download — it is a standard installer and needs no admin rights. Full walkthrough: Ollama on Windows setup guide.
Pull a text model and a vision model
The text model answers questions. The vision model is what lets it look at your screen.
ollama pull llama3.2:3b
ollama pull qwen2.5vl:3bRoughly 5 GB combined. Do this once while you have a connection — after that you are free of the network.
Install something to drive it
Ollama by itself is a terminal prompt. To get an assistant that watches your screen and talks back, you need an app on top. Clicky is one — it is free, open source, Windows-native, and finds an existing Ollama install on its own.
Confirm nothing is phoning home
Do not take anyone's word for it, including ours. Disconnect your wifi and ask a question. If you get an answer, the whole pipeline really is running on your PC.
Keeping it fully private
Running a local model is necessary for privacy but not automatically sufficient — an app can still reach the network for other things. Two settings worth checking in whatever you use:
- Turn web search off. It sends your question to a search engine, which defeats the point if privacy is why you are here.
- Use local speech, not cloud speech. Cloud transcription and cloud voices upload your audio.
In Clicky both live in the tray menu, and it ships with local speech recognition by default. There is also a privacy guard that skips screen capture on sensitive pages such as banking and login screens.
Is it worth it?
If you want the strongest possible answers, a cloud model is still better and you should use one. If you want an assistant that costs nothing, works on a plane, and never sends a screenshot of your work to a company, local is genuinely good enough now — which was not true a couple of years ago.
The pragmatic setup is both: local by default, with a cloud key you can switch to for the hard questions. Clicky supports Claude, OpenAI, Gemini and GitHub Copilot alongside Ollama, switchable from the tray without a restart.
Try it yourself — Clicky is free
A free AI assistant for Windows that reads your screen and answers out loud. No account, no API key, and it runs offline with Ollama.
↓ Download for Windows — freeWindows 10 / 11 · No sign-up