Your own LLM on your own computer: SI chat and agent on Ollama

What does a local LLM change?

Where your words go. With a cloud model, each question, the chat so far and the pages you add go to the provider you picked. With a local model the same request goes to Ollama, a program on your own computer, at the address 127.0.0.1, and the privacy audit’s line for it is short: it never leaves the computer. No key and no credits are involved, and the chat is unlimited.

Minda Super Intelligence is not a browser with a chat box added: it is the personal AI agent app for your computer, with a built-in browser the SI agent works in. A local model can run both, the SI chat that answers and, when the model can call tools, the agent that reads, clicks, types and fills forms in your signed-in workspaces.

How do I set it up?

1. Install Ollama on the same computer as Minda. Until it runs, Settings → AI → On this computer shows Ollama as Not available, with a Get Ollama link.
2. Download a model in Ollama. For the SI agent, pick one that calls tools; for screenshots and pictures, one that sees pictures. Any model can chat.
3. Back in Settings → AI → On this computer, the row now reads Running with the number of models. Press Check again if it does not.
4. In the chat, open the model menu next to Send and pick your model under Ollama (on this computer) · free. With no model picked, a chat starts on a local model before any cloud one.
5. To be sure nothing goes to a cloud model by mistake, turn on Keep SI on this computer in Settings → AI → Privacy (next section).
6. Ask something, then open the AI activity log (Settings → AI → Privacy → AI activity): the request lists the host 127.0.0.1:11434, which is Ollama on your computer, and the log counts the last 30 days’ requests as cloud or local.

Is picking a local model the same as Keep SI on this computer?

No, and the difference matters. Picking a local model puts that chat’s model on your computer, but the agent’s other tools still reach their services when they run: web search sends its words to a search engine, translation sends the text to a translation service, and the calendar and Notion connectors read from theirs.

Keep SI on this computer closes all of it at once. Every cloud provider, Minda’s hosted models, cloud speech and the agent’s search, translation, connector and picture tools are refused before anything is sent, and not even a model list is fetched from a cloud provider while it is on. The model menu lists your local models and shows each cloud provider as "cloud (off)"; picking one of those turns the setting off.

What can the SI agent do on a local model?

As much as the model can. Minda reads each model’s abilities from Ollama, whether it calls tools and whether it sees pictures, and the agent works within them.

• Read and act on pages, with a model that calls tools: it opens, reads, clicks, types and fills forms in your workspaces, with Ask before paying and the action log as with any model. A model without tools just chats.
• Look at your screen, after your OK each time: a model that sees pictures gets the picture; one that does not gets the text Windows reads in it, on your computer (Windows only).
• Minda’s own calendar: its tools work with a local model, since the calendar is kept on your computer and they send nothing anywhere.
• Compare models: two or three local models can answer one question side by side, and you keep the answer you like.
• Reply drafts and voice: with Keep SI on this computer, reply drafts use only a model in Ollama, and the voice assistant hears with Windows’ own recognizer and answers with the smallest model in Ollama that can do the job, so a spoken command gets its answer in seconds.

Which local model should I pick?

The one your computer runs well that can do the job. Settings says it plainly: local models are free, and slower without a strong GPU. Five things to weigh:

• Size. A larger model needs more memory and answers more slowly without a strong GPU. That is why, with Keep SI on this computer, spoken commands go to the smallest model in Ollama, while the chat uses the one you pick.
• Tools, for the agent. Only a model that calls tools can act on pages.
• Pictures, for screenshots. A model that sees pictures gets the screenshot itself.
• Context size. Minda uses each model’s own context window as Ollama reports it, up to 65,536 tokens. One with a small window, such as 32,768 tokens, can lose the earlier turns of a chat in which every agent tool is offered, so a follow-up arrives without what came before.
• Memory. Ollama loads each model into memory; two models compared side by side on a computer short of room answer one after the other.

What still leaves the computer?

The pages you and the agent open, and the features that are cloud by nature. A local model changes where the AI request goes, not how the web works: a site you open sees what any browser shows it. Minda’s own background requests carry on as the privacy audit lists them; an update check, for one, sends the version, channel, system, processor and package type, with no cookie, account or install id.

Minda’s servers receive browsing data only through features you turn on or use: sync, a team, community skills, or Minda’s hosted models when you pick one. Keep SI on this computer covers the AI; the audit covers the rest, feature by feature.

Can an organization require local AI?

Yes. An administrator can set the AiLocalOnly policy, through Windows Group Policy or Intune, or a policies.json on Windows, macOS and Linux, which turns Keep SI on this computer on for everyone on that computer and greys out the cloud models in the menu. AiProvidersAllowed can allow only Ollama instead. The admins of a team with paid seats can set the same policies for their members; a policy set on the computer itself wins.

For the wider list of what to ask of any AI tool in your browser, see the security checklist for browser AI tools.