Privacy
Corollary has no account, no analytics and no tracking. Your papers stay on your Mac unless you choose an AI model that runs somewhere else.
When Corollary goes online
| When | Where | What’s sent |
|---|---|---|
| Corollary opens | github.com | A check for a newer version. If there is one, it downloads from there. |
| You install TeX from Setup | The TeX Live archive (ftp.math.utah.edu, or ftp.tu-chemnitz.de as a backup) | Requests for the installer and package files |
| A paper needs a package you don’t have | The same server | Requests for that package. Finding which package you need happens on your Mac. |
| You turn on equations from photos | huggingface.co | One 1.2 GB download of the model |
| You choose an AI model on another computer or a cloud service | That server | The text you selected; for Write, the text around the cursor and your citation keys, labels and package names; or a compile error with the lines around it. The app shows “leaves this Mac” while one is selected. |
Nothing else uses the network. With Apple’s on-device model or Ollama on your Mac, AI edits work with Wi-Fi off.
What stays on your Mac
- Your projects, in the folders you keep them in. Build files go to a cache folder, not into your project.
- Snapshots saved before each AI edit, so you can go back. They’re in
~/Library/Application Support/io.github.calvinyz.corollary/checkpoints. - API keys you enter, in your macOS Keychain.
- A list of AI edits (when, which model, whether you kept it), used to draft an acknowledgments note. It contains none of your text.
- TeX, in
~/Library/Corollary/texlive, and the photo model, in~/Library/Application Support/io.github.calvinyz.corollary/models.
The models
Apple’s on-device model is part of macOS. The photo-to-LaTeX model, GLM-OCR, runs on your Mac with MLX. If you connect a cloud service with your own key, that service’s privacy policy covers what you send it.
This website
No cookies and no analytics. The page asks GitHub for the repo’s star count, so GitHub sees that request.
Corollary is open source, so you can read the code and check all of this.