Loci

Private AI for Mac

What makes AI private on a Mac?

Last updated:

A private Mac assistant runs a model where your work already lives: on your own computer, without sending every prompt and file to someone else's data center.

What to look for

  • The model should run on the Mac, not behind a privacy-themed interface connected to a cloud API.
  • Files and images should stay local when a local model processes them.
  • Online features should be optional and named at the moment they are used.
  • Model recommendations should account for memory, storage, and system responsiveness.

Verify local inference yourself

This test separates an AI interface that merely looks local from one that can actually answer without a hosted model.

  1. Open a local model in Loci

    Choose a model stored on the Mac or a supported local model service.

  2. Give it something specific

    Paste a paragraph, attach a local file, or ask for a structured outline.

  3. Disconnect the Mac

    Turn off Wi-Fi and unplug Ethernet. Keep the conversation open.

  4. Ask a follow-up

    Ask Loci to revise, compare, or extract something from the same material. A reply confirms that the AI path is local.

Follow the data, not the label

“Private,” “secure,” and “local-first” can describe very different products. The clearest way to judge a Mac AI app is to trace each action independently.

What leaves the Mac when using Loci
Action Sent to a data center/server? What actually happens
Local chat No Loci reads the prompt and generates the reply on your Mac. The prompt and response never go to a Loci server.
Local file or image analysis No Loci processes supported files and images with the model on your Mac. The content is not uploaded to a Loci server.
Optional web search Yes — query only The specific query goes directly from the Mac to the services selected for that search; with your permission, an Apple Maps nearby search can also use your device location as its search region. The rest of the conversation is not included, and none of these requests passes through a Loci server. Matching weather, air-quality, places, finance, holiday, sports, entertainment, or trending questions can first use structured services such as Open-Meteo, Apple Maps, Yahoo Finance, CoinGecko, Stock Analysis, Hargreaves Lansdown, Nager.Date, ESPN, JustWatch, Google Trends, Hacker News, or Wikimedia. General search in the quality source setting can use Bing, DuckDuckGo HTML, DuckDuckGo Lite, Mojeek, DuckDuckGo Instant Answer, and Wikipedia as direct search or fallback sources. If that direct roster finds no usable result, an eligible device can make one Google Search attempt through an on-device WebKit view, subject to availability and cooldown gates. The privacy source setting removes Bing from the direct roster and disables the Google Search rescue and the structured trending route, including Google Trends and its Hacker News and Wikimedia fallbacks; other matching structured services can still receive the query. Availability, cooldowns, and fallbacks mean not every query contacts every provider.
Model download File request The Mac retrieves model weights from a public repository. No conversation content is part of the download.

Why your Mac is a strong place to run AI

A Mac has more room for sustained work than a phone: a larger display, desktop files, a keyboard, and — in many models — enough unified memory to run capable local language models. That makes local AI useful for work that already lives on the computer:

  • Drafting and revising documents without uploading the source text to an AI provider.
  • Summarizing files you choose to attach.
  • Working through code, logs, plans, and structured notes.
  • Asking questions about supported images.
  • Using voice for a hands-free working session.

Model fit matters more than model size

Local models use the same pool of memory as macOS and every other open app. A model that technically fits can still leave too little headroom for the rest of the system.

Loci's custom hardware-sizing algorithm considers factors including RAM, processor generation, and the performance profile you want before recommending models for your Mac. The practical goal is not to load the biggest model name available. It is to choose the strongest model that stays responsive during the work you actually do.

Local AI is a good fit when

  • The material is private or commercially sensitive.
  • The task is writing, summarizing, planning, coding, or analysis.
  • You want the tool available without depending on a provider's uptime.
  • You prefer choosing and keeping models on your own computer.

Cloud AI may still be stronger when

  • The task needs frontier-scale reasoning or very large context.
  • The answer depends on extensive live research.
  • The Mac does not have enough free memory for an appropriate model.
  • Maximum capability matters more than local control.

Private browsing is not private inference

A private window keeps a page out of your history. It does not keep a prompt on the Mac: if the AI runs in the cloud, the request still has to reach the provider's server. Local inference changes the route itself, and there is no remote AI endpoint in it. More on why that boundary matters.

Questions about private AI on Mac

What makes an AI app private on a Mac?

The decisive question is where inference happens. In Loci, local models process prompts and generate replies on the Mac instead of sending the conversation to a hosted AI model.

Can Loci work with files and images locally?

Yes. Loci can work with files and images in chat. The selected model must support the type of input, and enough memory must be available to run it.

Does Loci for Mac need an account?

No account is required for local AI. You can install a compatible model and use the local workspace without creating a Loci identity.

Does private AI mean the Mac never connects to the internet?

No. Model downloads, app updates, optional web search, and optional diagnostics can use a connection. Local inference remains separate from those network features.

Will a larger model always be better?

No. A model that leaves too little free memory can slow the Mac or fail to load, so Loci recommends sizes that fit your hardware.

Which Macs can run Loci?

The Mac app requires macOS 14 or later. Actual model choices depend on the Mac's processor, unified memory, free storage, and other applications using memory at the same time.

Continue reading

Keep the work local

Bring the model to your Mac.

Use local models for writing, files, images, planning, and conversation without routing ordinary chat through a hosted AI provider.

Download Loci for Mac on the Mac App Store