1. Where does the model run?

This is the question the label is supposed to answer, and the one it most often blurs. An app can store everything on your disk, require no account, and still send each request to a remote API.

An AI product usually runs several kinds of models. Speech recognition turns audio into text. An embedding model turns text or images into numbers that make related content easy to find; Google's introduction to embeddings explains how. A generative model writes new text from what you give it. All of them see your data, so all of them count.

In Lumen, all three run on the Mac: Whisper large-v3-turbo, EmbeddingGemma 2, and Gemma 4 E2B. Plain keyword search doesn't need a model at all. If a model is missing or fails, Lumen shows an error instead of switching to a cloud service.

2. What can it read?

An agent needs access to your information to be useful, and that access should have a boundary you set.

Lumen reads only the folders you choose. Choosing one project folder doesn't give it the rest of your disk, and it doesn't change your files unless you review and apply a routine. It records only when you start a recording, or when you've turned on automatic recording, which is off by default. Importing an existing recording never turns on the microphone. Activity memory, such as which apps you used, is off until you turn it on, one source at a time.

Processing creates new data too: transcripts, extracted text, and a search index. Those are as sensitive as the originals. Lumen keeps them in its local storage on your Mac.

3. Where do the results go?

A locally generated summary is still a document that can be shared, synced, or backed up. Local processing only helps if the output stays put until you decide otherwise.

In Lumen, notes stay in the app until you export them. An export is a Markdown file you review and save to a folder you pick. If that folder is synced by another service, the file follows that service's rules from then on.

4. What protects it at rest?

Local isn't the same as encrypted. Data on your disk is as safe as the disk, the apps that can read it, and wherever your backups go.

Lumen doesn't add its own encryption layer, and it hasn't had an independent security audit. It relies on macOS, so turn on FileVault and check what your backups include.

Check the claim

Whatever product you're evaluating, ask how its “local” claim was tested. For Lumen, we've run transcription, note generation, and search with networking blocked by the operating system, and the app and its AI processes have no network access. The only network request is the one-time model download from Hugging Face during setup.

Those tests show where the processing happens. They don't show how accurate it is. Our evaluations so far are small, use synthetic inputs, and ran on one M1 Pro with 16 GB of memory.

This website is separate from the app and is a normal hosted site. Website privacy explains what the host can see.