Can you run AI jobs on your own computer for free?
Yes. Ollama is a free program that runs open AI models on your own computer, and Magic Folder turns it into folders that do real jobs. There's no subscription and no API key, and your files never leave your machine.
Every job comes back with a signed receipt, and receipts for local jobs say exactly that: "Free (ran on this computer)".
What is Magic Folder?
A folder with a subfolder for each job. It comes with six: Write product descriptions, Fix typos, Meeting notes to tasks, Image alt text and SEO, Draft a client email, and Ask anything (drop in files plus a note saying what you want).
You drop a file into a job folder and leave it. A small "working" note appears while it runs, then the result shows up in Done, with your original untouched and a receipt you can open in any browser. You can make your own jobs by copying a folder and editing one text file.
What do you need?
A Mac, Windows or Linux computer, Node.js 22 or 24, and Ollama with a model. Pick Node 24 specifically: the "LTS" download on nodejs.org is now Node 26, which Magic Folder's database driver doesn't support yet.
For the model, qwen3:8b is a good start. It's about a 5 GB download and runs comfortably on most recent computers; a machine with more memory will run it faster.
How do you set it up?
Five steps: install Node 24, install Ollama and pull a model, install Magic Folder with npm, create your folder with magic-folder init, and start it with magic-folder run using the Ollama engine. magic-folder doctor checks that everything is in place. The exact commands for each system are in the install guide, with a copy button on every step.
Once it works, you can have it start automatically when you log in, so the folder is always on. The guide covers that for Mac, Windows and Linux too, along with how to stop or remove it.
What do local models do well, and where do they struggle?
They're good at focused jobs: fixing typos, turning meeting notes into a task list, short product descriptions. For Fix typos, the list of changes isn't written by the AI at all; Magic Folder compares your original with the corrected copy and lists every real difference, so it can't invent one.
They're smaller than Claude, so writing comes out plainer, and now and then a job doesn't finish. When that happens the receipt says so, nothing is charged, and you can drop the file in again. Images only work with a model that can see them, and PDFs or Word files need saving as .txt or .md first.
When should you use Claude instead?
For client-facing writing, longer jobs, or anything with images, Claude does noticeably better work. You can keep Ollama as the default and switch individual jobs to Claude with one line in that job's settings, using your own Anthropic API key.
Two differences to know: Claude's receipts show the estimated cost of each job, and the job's files are sent to Anthropic's API instead of staying on your computer.
Is my data private?
With Ollama, yes: the AI runs on your machine and the files stay there. Magic Folder fingerprints the files itself and keeps receipts and its signing key in a folder on your computer. Even checking a receipt happens in the browser, with nothing uploaded.
It's free and open source. Start with the install guide, or read how signed receipts work.