Private personal agent · In development

Your personal agent.
Entirely local.

Understand recordings. Find information. Create useful work. Lumen is being built to turn conversations, files, and ideas you choose into useful work, with AI models running on your device.

01Understand recordings and conversations.
02Find information in what you select.
03Create useful work you can review.

A labeled walkthrough

From a conversation to a useful next step.

This is the workflow Lumen is being built around. Every step below is sample material. This website does not record you, accept your files, or run the product’s AI.

Sample · synthetic · not your data. The creation step is a concept written for this page, not a result from Lumen.

Sample session

Team catch-up

Example notes and transcript. No audio was recorded, and no model ran.

Summary. The team kept the timeline and planned a layout review before adding work. Source 00:12

Next steps. Send the project update — Speaker 2, Friday. Prepare the layout for review — Speaker 1, Monday.

In the macOS app this is the built-in sample meeting. It shows notes, source times, and export. It is not a recording of a real conversation.

The requirement

The models run locally too.

Local means the speech, embedding, and generation models run on the supported device. Personal recordings, prompts, indexes, and outputs stay in local storage by default. A missing or failed local model is a local failure. It does not switch to a cloud model.

Chosen input → On-device processing → Reviewed output

Established, and not yet

  • The signed macOS app and its helper have no network entitlements. Search checks have run with networking denied.
  • Adapter contracts for transcription, summaries, and embeddings exist. A synthetic local run is recorded in the project. It is not a quality or hardware benchmark.
  • There is no application-specific encryption layer in this slice. Device protection and backups are separate.
  • Nothing here downloads model weights for you, and nothing sends the work to a remote completion API.

Direction and readiness

The aim is broad. The preview is specific.

Lumen is intended across devices and operating systems. Each supported device would run its own models. The software that exists today is a macOS preview.

Recordings and meetings In the preview

Manual meeting capture, rough notes, editable actions, and recovery are implemented. The sample meeting is labeled and does not record the room. Live microphone and call-audio behavior still need validation.

Chosen-information search In the preview

Keyword and phrase search works on text and text-bearing PDFs in folders you select, with passages and page-aware PDF viewing. Image files can be indexed and previewed. Visual description search needs compatible local tools and has no quality claim.

Creation In development

Editable notes and reviewed Markdown export are the current foundation. Recaps, briefs, plans, and other formats are the direction. The sample recap on this page is a concept.

Scoped local tools Planned

Future help with a narrow local routine would need preview, permission, cancellation, and recovery. The current search workflow does not change your source files. An export is a new file you choose to save.

Other platforms Planned

Windows, Linux, and mobile are candidates, not releases. No sync service moves work between devices. Optional device-to-device transfer would be a separate design.

Activity memory Off

Activity memory is off and has no capture implementation. A personal agent, here, is help with context you choose. It is not continuous observation.

What can be shown

Two screens from the signed macOS app.

Captured 15 October 2026 from the signed development bundle, using synthetic evaluation files and the built-in sample meeting. They are not a hardware or model benchmark. No public installer is offered.

Lumen on macOS showing three keyword results for INV-2048 from synthetic office files. Visual search is marked not prepared. Activity memory is off.
Keyword search for INV-2048 in the synthetic fixture folder. Visual search is marked “Not prepared.” Activity memory is off.
Lumen’s sample meeting, Team catch-up, with a banner stating that no audio was recorded and no local inference ran.
The built-in sample meeting. The app says no audio was recorded and no local inference ran.

FAQ

Plain answers

Do the models run locally?

Yes. That is a requirement of the intended agent, covering speech, embeddings, and generation. Current real-model quality is not an established public result. There is no remote-completion fallback and no automatic model download.

What is available, and on which platforms?

The current implementation is a macOS preview for Apple Silicon development, with macOS 14 as the deployment target. It has been checked on an M1 Pro with 16 GB of memory. That is one development machine, not a compatibility list. Windows, Linux, and mobile versions are not available. There is no public download or installer on this site. The preview is built from source.

Does it record me or watch what I do?

No. Activity memory is off and unimplemented. In the app, recording and folder access are deliberate actions. This website does not record the visitor, does not ask for a microphone, and does not upload files.

What can it create today?

The preview supports editable notes, action text, and a Markdown export you review before saving. Drafts, briefs, plans, and wider formats are planned. The recap in the sample section is labeled as a concept because the product did not generate it.

Does local mean encrypted or audited?

No. This slice has no application-specific encryption layer, and no independent security audit is claimed. Local processing means where the work runs and where the product stores it. Device protection, backups, and any folder you export into are separate.

Will my devices sync?

No sync service is implemented. Optional transfer between your own paired devices is a future idea, separate from running on more than one operating system. Each supported device is intended to do its own inference.