Guide
Local AI means running an AI model on your own computer instead of a company's servers, so your prompts and files don't have to leave your Mac. It's worth it when you have private material and a narrow job you repeat often. Ollama fits people who want an open-source runner with a local API. LM Studio fits people who want a desktop app. Dayflow can use either to analyze your screen on-device.
You have something open that matters: a contract, a page of your journal, a spreadsheet with client names in it. A chat window sits next to it. You select the text and start to paste, then stop. Where does this go once you hit enter, and who keeps it? You close the tab and do the job by hand, or you paste it anyway and feel a little uneasy for the rest of the afternoon.
Then someone mentions “local AI” as if it settles all of that. Run the model on your own Mac, they say, and nothing leaves. It sounds right, and also vague. Which model, what program, will your laptop run hot, and is it worth an evening of setup?
This guide explains what local AI is in plain terms, why it matters where a model runs, and how to decide if it’s worth it for you. Most of the deciding happens with a notepad and a look at your Mac’s settings. It also covers the two common ways to run models on a Mac, what they cost, and one specific use: a private record of your working day. Dayflow, one of the tools mentioned, is our product, so weigh that section accordingly. Product details come from each vendor’s own documentation as checked in early October 2026. Nothing here comes from hands-on benchmarks.
When you type into a chatbot in your browser, your words travel to the provider’s servers. A model there produces an answer and sends it back. What happens to your text afterward, whether it’s stored, for how long, and who can see it, depends on that company’s terms and your settings. Most of the time that’s fine. Sometimes it isn’t, and you can’t always tell from the chat window.
With local AI, the model is a file on your disk and your Mac’s chip does the computing. Nothing has to leave the machine for you to get an answer. That brings a few real benefits: it works offline once the model is downloaded, there’s no per-question bill for local use, and sensitive material stays where it already is.
The costs are just as real, and they’re the ones people tend to discover late:
The word “AI” also blurs three separate things: the model (the file that does the thinking), the app you talk to, and the company behind it. Local AI changes where the model runs. It doesn’t automatically change the other two. Keeping those apart makes every decision below easier.
What to do: For a few working days, keep a sheet of paper next to your keyboard with three columns: “Fine anywhere,” “Fine with a company I trust,” and “Never leaves my machine.” Each time you use an AI tool, or decide not to, write down a word or two about the material in the right column.
An illustration: A freelance designer’s sheet might read: “rewording a portfolio blurb” under fine anywhere, “summarizing a client brief” under trusted company, and “notes from a therapy-adjacent coaching call” and “unreleased client logos” under never leaves.
How to tell after a week: Look at the third column. If it’s empty, local AI probably isn’t worth the setup yet, and your energy is better spent on the cloud tool’s privacy settings. If it’s long, and the same kinds of material keep showing up, you’ve found a real reason to run something locally.
What to do: From your third column, pick the single most repeated task and write it as one sentence: the input, what you want back, and how often. Local models tend to fit narrow, repeatable jobs like summarizing, rewriting, tagging or sorting better than “know everything” questions.
An illustration: “Every Friday, turn my rough meeting notes from the week into a five-line summary for my own records.” That’s a clear input, a clear output and a fixed rhythm. Compare it with “help me with work stuff,” which no model, local or not, can be judged against.
How to tell after a week: Did you do this job by hand at least once and wish you didn’t have to? If yes, you have something concrete to test. If the job only came up once, pick another one or wait until a real need shows up.
What to do: Open the Apple menu, choose About This Mac, and write down three things: the chip (Apple M-series or Intel), the memory, and the macOS version. Then check free space under System Settings → General → Storage. Keep the note. You’ll compare it with the requirements in the next sections.
An illustration: A note might say “M-series, 8GB memory, macOS 14, about 40GB free.” That person can install both common runners, should stick to smaller models, and may need to clear space first.
A few hard limits from the vendors’ own pages, so you can rule things out early:
How to tell it’s working: You can answer “which runner can I use, and roughly how much room do I have?” without looking anything up. If the answer is “neither” (an Intel Mac and no wish to run CPU-only), you’ve just saved yourself an evening.
What to do: Before adding a new tool, open the privacy or data settings of the AI service you already use. On paper, answer three questions: Are my chats stored? Are they used for training? Can I turn either off?
An illustration: Someone might find that their chatbot has a setting to exclude chats from training, and that turning it off covers everything in their middle column. Only the “never leaves” pile still needs a local answer.
How to tell after a week: If adjusting settings made you comfortable with most of what you wanted to paste, local AI becomes a small tool for a short list of material, not a replacement for everything. That’s a perfectly good outcome.
These are the two runners most people start with on a Mac. Both run open models on your own machine, and both have paid plans, but those plans pay for cloud features, not for running models locally.
| Ollama | LM Studio | |
|---|---|---|
| What it is | Runs open models through a CLI, desktop app and local API, with optional cloud models | Desktop app for running open models with llama.cpp, plus Apple’s MLX on Apple silicon |
| Mac support | macOS 14 or newer; Intel Macs CPU only | Apple silicon only, macOS 14.0 or newer |
| Other platforms | Windows, Linux, Docker | Windows (x64 and ARM), Linux (x64 and ARM64) |
| License | Open source (MIT) | Proprietary; free for home and work use |
| Local use | Free | Free |
| Paid plans | Pro $20 / mo. or $200/yr billed annually; Max $100 / mo.; Team $500 / mo. (early access); Enterprise custom | Bionic+ $20USD / month; Pro $100USD / month |
| Free trial | None stated | None stated |
| Data when local | Vendor says it doesn’t see your prompts or data when you run locally; cloud features can be disabled with OLLAMA_NO_CLOUD=1 | Docs say it can run entirely offline once model files are downloaded; paid cloud inference is US-hosted and described as Zero Data Retention |
Ollama suits you if you want open-source software, plan to connect other apps through its local API, or own an Intel Mac. LM Studio suits you if you prefer a desktop app and have Apple silicon. Either way, pick a model using the runner’s own guidance for your memory. This guide doesn’t recommend specific models or sizes, because those change often and the vendors’ docs are the right source.
What to do: Install one runner, download one model, and use it only for the job you named earlier. Keep doing that job the old way too, at least once, so you have something to compare against.
An illustration: On Monday, the designer pastes a client brief into the local model and asks for a five-line summary, then writes their own summary for comparison. By Wednesday they’ve stopped writing their own because the model’s version is close enough.
How to tell after a week: Ask yourself three questions. Did you use it more than twice? Was the output good enough that you didn’t redo it? Did the fan, battery or disk space bother you? Two yeses and a no on the last one means keep it. Otherwise, uninstall it and delete the model files (Ollama keeps them in ~/.ollama/models on a Mac) to get your space back.
What to do: Write a one-line rule on a sticky note: “Pile three goes local, everything else can go to the cloud.” Stick it on the edge of your screen. Local AI rarely needs to do everything, and forcing it to usually means slower answers on jobs that were never sensitive.
An illustration: Brainstorming a newsletter headline goes to the cloud chatbot. Summarizing a private coaching note goes to the local model. Neither choice takes any thought once the rule exists.
How to tell after a week: Look back at the week. Did you ever paste pile-three material into a cloud tool out of habit? If not, the rule works. If you did, move the local tool somewhere you’ll see it, like the Dock, so it’s the easy option.
Your screen is about as pile-three as material gets. It shows every message, document and tab you touched. It’s also what you’d need if you wanted AI to answer “what did I actually do today?” without starting a timer or reconstructing the day from memory.
Dayflow is a free, open-source Mac app that records the screen at one frame every 10 seconds and has AI write the day into a readable timeline. Full recordings and the timeline are stored on your Mac. In its Local AI mode, a supported model runs on your Mac through either of the runners above, and screenshot analysis stays on-device. Setup and local inference use your own hardware.
There are two other modes. You can point it at your own cloud provider, where selected screenshots go to that provider under its terms, and the free app doesn’t include paid API usage. Or you can use Dayflow Pro at $20 a month, which manages cloud analysis. Under the privacy policy, Pro deletes screenshots after processing and may retain Activity Data for sync. The local and cloud processing page spells out each boundary. The app itself is $0, MIT licensed and needs macOS 14 or later, and you can export a date range as Markdown from Settings → Export.
Dayflow is our product, and we publish this article.
Local AI moves the model onto your own Mac, so prompts and files don’t have to leave it. That’s valuable when you have material you’d never paste into a cloud tool and a narrow job you repeat often. It’s less valuable as a general replacement for cloud AI, because your hardware does all the work and laptop-sized models have limits.
Start on paper. Sort what you’d paste, name one job and check your Mac. Only then install a runner: Ollama if you want open source and an API, LM Studio if you want a desktop app on Apple silicon. Give it a week on that one job, keep a cloud tool for everything else, and let the result decide. If the private thing you most want help with is remembering your own working day, that’s the one place Dayflow fits.
Local AI is an AI model that runs on your own computer. The model file sits on your disk and your Mac's chip does the work, so a prompt doesn't have to travel to a company's server to get an answer. A cloud chatbot does the opposite: your words go to the provider, get processed there, and come back. The trade-off is privacy and offline use versus your own hardware doing the heavy lifting.
It can be, as long as you run a local model. Ollama's documentation says it doesn't see your prompts or data when you run locally. LM Studio's docs say it can operate entirely offline once model files are downloaded. Both also offer optional cloud models, and those don't stay on your machine. Ollama lets you switch cloud features off with OLLAMA_NO_CLOUD=1. Check which mode you're in before pasting anything sensitive.
Start with About This Mac. Ollama needs macOS Sonoma (14) or newer. Apple M-series Macs get CPU and GPU support, and Intel Macs run on CPU only. LM Studio needs Apple silicon and macOS 14.0 or newer, with 16GB or more of RAM recommended. Smaller models may work on 8GB machines. Ollama doesn't publish a RAM minimum, so check its docs for your model. Free disk space matters too.
Running models locally with Ollama is free, and the software is MIT licensed. Paid plans are for cloud-hosted models and higher cloud usage. Pro is $20 / mo. or $200/yr billed annually, Max is $100 / mo., Team is $500 / mo. in early access, and Enterprise is custom. The free plan includes starter cloud usage credits and starter cloud models. The pricing page lists no trial.
Yes. LM Studio announced in July 2025 that it's free to use at home and at work, with no form to fill in. The app is proprietary, not open source, and its terms license it for personal and internal business use. The free plan covers running local models with llama.cpp and MLX. Bionic+ at $20USD / month and Pro at $100USD / month add cloud inference and web search.
It depends on how you like to work. LM Studio is a desktop app, supports Apple's MLX on Apple silicon alongside llama.cpp, and is free for work but not open source. Ollama has a CLI, a desktop app and a local API with Python and JavaScript libraries. It's MIT licensed and runs on Intel Macs in CPU-only mode. If you plan to connect other apps, Ollama's API is a natural fit.
Once the model is on your disk, yes. LM Studio's documentation says it can operate entirely offline once model files are downloaded. Ollama runs local models on your machine, and its cloud features can be turned off completely. You need a connection to download the runner and the model files the first time. Those files can be large, so plan the download for a good connection.
No. Dayflow has three ways to analyze screenshots. Local AI uses a supported model through Ollama or LM Studio, and the analysis stays on your Mac. Your own provider sends selected screenshots to the cloud service you configure, under that provider's terms. Dayflow Pro, at $20 a month, manages cloud analysis for you. The free app costs $0 either way and stores recordings and the timeline on your Mac.
No. LM Studio or Ollama beats Dayflow here. Dayflow is a Mac app that records your screen and writes your day into a timeline. It isn't a general chat or writing assistant. If you want to ask a model questions, draft text or write code privately, install LM Studio for a desktop app or Ollama for an open-source runner with a local API, and use it directly.
Often not for broad, open-ended questions. Many people find that models small enough to run on a laptop handle narrow jobs well, like summarizing, rewriting or sorting, but fall behind the largest hosted models on hard reasoning and general knowledge. Speed also depends on your hardware. The way to find out is to try one specific job of yours for a week and judge the output yourself.
Dayflow is $0 and open source. Full recordings stay on your Mac, with local AI available for on-device analysis.
Download for MacOS macOS 14+ · MIT licensed · no account required