Guide
Yes, for focused jobs on hardware that fits. Ollama (free, MIT licensed, macOS 14+, GPU support on Apple M series) suits people who want a command line or local API. LM Studio (free for home and work, Apple silicon and macOS 14+, 16GB+ RAM recommended) suits people who want a desktop app with MLX. Dayflow can use either to analyze screenshots on-device. Both runners also sell optional cloud plans.
You have a document open that you would rather nobody else read. Maybe it is a client contract, a draft performance review, or a page of notes from a doctor’s appointment. You want an AI to summarize it, and your cursor hovers over a chat box run by a company whose servers you will never see.
So you search for something that runs on your own Mac. Within minutes you are looking at model names with numbers in them, downloads that would eat a big slice of your disk, and apps that say “local” on the home page and “cloud credits” on the pricing page. It is hard to tell what is real.
This guide covers what “never leaves your Mac” actually means, how to check whether your machine can do it, and some low-effort ways to try it without fooling yourself. It sticks to two local model runners, Ollama and LM Studio, plus Dayflow, which is our product and comes up near the end. Every claim about these tools comes from comparing vendor documentation: the Ollama and LM Studio pages as checked on October 4, 2026, and Dayflow’s published facts for version 2.2.0. Nothing here comes from hands-on benchmarks.
Part of the confusion is the word “local.” It gets used for two different things: where your files are stored, and where the processing happens. An app can keep everything on your disk and still send pieces of it somewhere else to be analyzed. When you read “local,” ask which of the two it means.
The second reason is that the tools are hybrid on purpose. Ollama’s free plan includes starter cloud usage credits and access to starter cloud models alongside local ones. LM Studio’s paid plans add optional US-hosted cloud inference, which the vendor describes as Zero Data Retention. None of that is a trick, and both vendors are clear about it. But it means “I’m using Ollama” or “I’m using LM Studio” does not tell you on its own whether a given answer came from your Mac or from a data center.
The third reason is hardware. A model is a large file that your Mac has to load and work through. Whether that is comfortable depends on your chip, your memory and your free space, and the vendors describe their requirements differently.
Last, models change quickly. Any article that tells you “the best model to run” is out of date soon after it is written. That is why this guide names no models, sizes or speeds. The model pages on each vendor’s site are the place to check those, on the day you download.
Before any download, take a sheet of paper and draw two columns. On the left, write the kinds of material you would never paste into a cloud tool. On the right, write the kinds you would not mind. Be specific: “client contracts,” not “work stuff.”
Illustration: a freelance designer might put client briefs under NDA, invoices with bank details and personal journal entries on the left, and blog post drafts and public product descriptions on the right.
This sounds too simple to matter, but it changes the question from “should I use private AI?” to “which of my tasks need it?” Many people find the left column is shorter than they expected, and that is useful to know.
How to tell after a week: look at the sheet again. If it changed a decision, even once, it is working. If almost everything sits on the left, local AI may deserve to be your default. If the left column is empty, you may not need local AI at all, and that is a fine outcome.
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 your free disk space in System Settings. No new software needed.
Compare what you wrote with what each vendor lists:
| Requirement | Ollama | LM Studio |
|---|---|---|
| macOS version | macOS Sonoma (v14) or newer | macOS 14.0 or newer |
| Apple silicon | CPU and GPU support (GPU via Metal) | Supported, with llama.cpp and Apple’s MLX |
| Intel Macs | CPU only | Not supported |
| Memory guidance | No RAM minimum stated in the docs; see the vendor’s docs for each model | 16GB+ RAM recommended; 8GB may work with smaller models |
| Disk | Models are tens to hundreds of GB; stored at ~/.ollama/models | See the vendor’s docs |
| License | Open source (MIT) | Proprietary; free for home and work |
| Cost for local use | $0 | $0 |
Disk is the one people forget. Ollama’s docs say models are tens to hundreds of GB in size, so a laptop that is already near full will run out of room before you have tried much.
Illustration: someone on an older Intel MacBook would find LM Studio is not an option and Ollama would run on the CPU only. Someone on a recent Apple silicon Mac with plenty of memory and a mostly empty drive can try either.
How to tell it is working: this check pays off before the week even starts. If you could rule a runner in or out from your notes without downloading anything, it did its job. If you are unsure, the vendor’s system requirements page is the next stop, not a forum thread.
Look at the left column of your list and choose one task you do every week that is small, repetitive and low-stakes. Write it on an index card and keep it by your desk. Resist the urge to test the model on everything at once.
Illustration: turning rough notes from a confidential client call into a tidy five-line summary. You already know what a good result looks like, so you can judge the output in seconds.
A dull task works because you are testing whether local AI fits your week, not whether it can do something impressive. A model running on a laptop does not have to match the biggest cloud models to be useful for a job like this. Treat every answer as a draft and read it before you rely on it.
How to tell after a week: did you reach for it more than once without forcing yourself? Was the output usable with light edits? If yes, add a second task. If you kept going back to doing it by hand, either the task or the setup is wrong, and it is better to know now.
Once your model has finished downloading, turn off Wi-Fi and unplug any network cable. Then run the task from your index card again.
This is the most convincing test available, and it needs no expertise. LM Studio’s docs state that it can operate entirely offline once model files are downloaded. Ollama’s docs say: “Ollama runs locally. We don’t see your prompts or data when you run locally.” For Ollama, you can also switch off its cloud features entirely by setting OLLAMA_NO_CLOUD=1, which is worth doing if you never want a cloud model picked by accident.
Illustration: you summarize the same meeting notes with the network off. The answer arrives, a little different in wording but just as usable. Now you know that this setup, with this model, did the work on your Mac.
How to tell after a week: you should have run your real task offline at least once. If anything failed or behaved differently with the network off, something in your setup depended on the cloud. Find out what before you trust it with the left column.
Because both runners offer cloud models next to local ones, it is easy to lose track. On a sticky note, write the runner, the model you chose, whether it is local or cloud, and whether you turned cloud features off. Stick it on the edge of your screen.
Illustration: “Ollama, local model, cloud off (OLLAMA_NO_CLOUD=1), set up Oct 10.” Or “LM Studio, local model, free plan, no cloud inference.”
It is also worth knowing what the paid plans are for, so they do not confuse you. Ollama’s Pro is $20 / mo. or $200/yr billed annually, Max is $100 / mo., Team is $500 / mo. (early access), and Enterprise is custom; the paid plans are metered in cloud usage credits. LM Studio’s Bionic+ is $20USD / month and Pro is $100USD / month, and those tiers pay for cloud inference and web search, not for running local models. If staying on-device is the whole point, the free plan of either is the private one. Neither vendor lists a trial; see each vendor’s pricing page for current details.
How to tell after a week: someone asks “where did that answer come from?” and you can say without opening the app. If you have to check, the note is not visible enough, or your setup has drifted.
Local processing protects your input up to the moment you copy the output somewhere else. If you summarize a private document on your Mac, then paste the summary into a cloud chat to “polish it,” the private part has left.
On the same sheet of paper, write where outputs from left-column tasks are allowed to go: a notes file on your Mac, a local folder, an email to the one person who should see it. Write where they are not allowed to go.
Illustration: the summary of the confidential client call goes into a local notes file and into an email to the client. It does not go into a shared document tool or a cloud assistant for rewording.
How to tell after a week: scroll back through where you pasted things this week. If every left-column output stayed inside the boundary you wrote, the habit has stuck. If one slipped, notice what pulled it out, usually a quick “make this sound better,” and decide whether that step can also run locally.
Local AI matters most when the material is something you would never upload, and few things are more personal than a record of your whole screen. Dayflow, a free, MIT licensed app for macOS 14 or later, records the screen at one frame every 10 seconds and has AI write the day into a readable timeline. You do not have to remember to start a timer or rebuild the day from memory when someone asks where your day went. The full recordings and the timeline are stored on your Mac.
The analysis is where everything above applies. In Local AI mode, a supported model runs on your Mac through Ollama or LM Studio and screenshot analysis stays on-device, using your own hardware. Alternatively, selected screenshots can go to a cloud provider you configure, under that provider’s terms, or to Dayflow Pro at $20 a month, which manages cloud analysis and, under the privacy policy, deletes screenshots after processing and may retain Activity Data for sync. One honest caveat: if you ask a cloud assistant about your day, the Dayflow data it retrieves goes into that conversation. Local storage does not make a cloud chat local. The three modes are laid out on local and cloud processing.
Dayflow publishes this article.
If any screen recording feels like too much, time tracking without surveillance covers lighter approaches. If battery is your worry, see what screen recording does to battery, and for the wider privacy questions, is Dayflow safe.
Private AI that stays on your Mac is real today, with conditions. You need a machine that meets the runner’s requirements, enough disk for files that can run to hundreds of GB, and the discipline to know which model you picked and where its output goes. Ollama suits people who want an MIT licensed runner with a command line and local API. LM Studio suits people on Apple silicon who want a desktop app with MLX support. Both are free for local use, and both sell cloud plans you can ignore if privacy is the point.
Start with paper, not downloads: list what must stay private, check your Mac, pick one dull task, and prove it works with the Wi-Fi off. If the private material you care about most is your own working day, that is where Dayflow’s local mode fits.
Yes, once the model files are on your disk. LM Studio's docs state it can operate entirely offline after models are downloaded. Ollama says it runs locally and does not see your prompts or data when you run locally, and its cloud features can be switched off with OLLAMA_NO_CLOUD=1. The simplest check is to turn off Wi-Fi and run the same task again.
Not when you run a model locally. The vendor states: 'Ollama runs locally. We don't see your prompts or data when you run locally.' Ollama also offers cloud-hosted models, and for those it says it processes prompts and responses to provide the service but does not store or log that content. Setting OLLAMA_NO_CLOUD=1 disables the cloud features entirely.
It depends on the runner and the model. LM Studio recommends 16GB or more of RAM on a Mac and says 8GB machines may work with smaller models. Ollama's macOS, GPU and FAQ docs state no RAM minimum, so check the page for the specific model you want before downloading it rather than relying on rules of thumb found elsewhere.
LM Studio does not. Its Mac version requires Apple silicon and macOS 14.0 or newer. Ollama needs macOS Sonoma (version 14) or newer and does run on Intel Macs, but x86 Macs get CPU support only, while Apple M series Macs get both CPU and GPU support. Check About This Mac to see which chip you have.
No. LM Studio is a proprietary desktop app, and its terms prohibit modifying, redistributing or reverse engineering the software. It is free to use at home and at work with no separate commercial license needed. If an open-source license matters to you, Ollama is MIT licensed, and so is Dayflow.
Cloud features, not local use. Ollama's Pro is $20 / mo. or $200/yr billed annually, Max is $100 / mo., Team is $500 / mo. (early access), and paid plans are metered in cloud usage credits. LM Studio's Bionic+ is $20USD / month and Pro is $100USD / month, paying for cloud inference and web search. Running local models is free on both.
A lot. Ollama's documentation notes that models are tens to hundreds of GB in size. On macOS, Ollama stores them in ~/.ollama/models, which is the folder to check if your disk suddenly fills up. Before downloading, look at the size listed for the specific model and compare it with the free space shown in your Mac's storage settings.
No. If you want to chat with a local model, summarize documents, or use one with coding tools, use LM Studio or Ollama directly. Dayflow uses those runners for one job: analyzing screenshots of your own screen to write a timeline of your day. It is a record of your work, not a general-purpose interface for local models.
In Local AI mode, a supported model runs on your Mac through Ollama or LM Studio and screenshot analysis stays on-device. Full screen recordings and the timeline are stored on your Mac in every mode. One caveat: if you ask a cloud assistant about your day, the Dayflow data it retrieves goes into that conversation, which is then not local.
Yes. Dayflow is $0 and MIT licensed, and it runs on macOS 14 or later. There is an optional Dayflow Pro plan at $20 a month that manages cloud analysis for you. The free app does not include paid API usage if you choose to connect your own cloud provider, and it does not provide invoices, payroll approvals, team reports or website blocking.
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