Lumen is in development. This page lists what works in the current macOS preview, what has been checked, and what still needs validation.
The preview runs on macOS 14 or later on Apple Silicon. There is no public installer: you build the signed app from source, as described in Getting started. It isn't a finished or cross-platform product.
Available in the current preview#
| Workflow | Current scope |
|---|---|
| Keyword search | Text and text-bearing PDFs in chosen folders; exact words and numbers; passages and source pages. No model needed. |
| Intelligent search (search package) | Semantic text and image retrieval, OCR, rich documents, audio and video moments, optional code search, filters, exclusions, and automatic folder refresh |
| Recordings | Microphone and call audio (call audio on by default once allowed), import of existing recordings, pause and stop controls, local audio, and recovery |
| Meeting notes (local AI bundle) | Local speech recognition, source-linked generated notes, editable rough notes and next steps, meeting-text search, and speaker labels within a meeting after recording stops |
| Meeting copilot | Briefs before meetings, commitments caught while recording, follow-up drafts you send yourself, and next steps sent to Reminders after you approve |
| Search or ask (⌘K) | One bar for files and meetings; Ask answers from your files, meetings, and memory with sources |
| Memory and Today | Opt-in sources (calendar, meetings, chosen folders, app activity, clipboard), each off by default; suggestions you accept; Today dashboard with time by app, category, and project; pause, exclusions, and delete |
| Routines | Inbox-folder extraction to CSV, Downloads cleanup, and after-meeting filing, each previewed before it runs, with Trash instead of delete and undo; Shortcuts actions |
| Export | Reviewed Markdown, with an optional transcript, saved to a destination you choose |
| Local control | Folder access only where you grant it, local storage, retention settings, and no account or remote-completion fallback. Models download once from Hugging Face at setup. |
Intelligent search and meeting notes need their optional local packages. For exact behavior and limits, see How local AI is prepared, Find files and information, and Meetings and recordings.
What has been checked#
Development checks cover real text extraction and SQLite retrieval, scope and cancellation behavior, synthetic audio-file capture, recovery and retention, and verified runtime contracts. Small synthetic offline checks also ran real speech recognition, generated notes, semantic and visual retrieval, and the expanded native search engine on one M1 Pro with 16 GB RAM.
The native UI has been exercised for fixture search, source previews, the labeled sample meeting, and reviewed export. Memory, Today, routines, the meeting copilot, permissions, and model download are covered by automated tests and views rendered with synthetic data; they haven't yet been tried by hand on other Macs. The documentation screenshots are those renders, not model-quality measurements.
These checks show that the workflows run on that setup. They don't establish broad language accuracy, codec coverage, universal relevance, a minimum RAM requirement, or independent security or accessibility certification. Developers can reproduce the evaluation workflows.
Validation still needed#
- Live microphone and call-audio routing, and permission denial or revocation, on each target device.
- Longer real meetings, overlapping speech, names and numbers, more languages, and note faithfulness.
- More Mac configurations, an 8 GB target, memory pressure, battery use, and slow storage.
- Full VoiceOver and macOS accessibility settings.
- Notarized distribution, a clean-device installation path, the real model download over slow networks, and stable upgrade behavior.
- The real macOS permission prompts for Calendar, Reminders, and Accessibility, Shortcuts actions in the Shortcuts app, and CPU and battery use with memory sources on.
Important: Automatic meeting mode uses microphone-activity heuristics. A running timer or a nudge doesn't prove that Lumen captured the call audio you intended, so always check the recording state and any visible warnings.
Still in development#
These features aren't available in this release:
- Dictation into other apps
- Live speaker labels during recording, speaker recognition across meetings, and verified speaker identity
- Screen content and browser history as memory sources
- Connections to other apps, such as Slack
- Windows and Linux apps, and device-to-device continuity
No sync service is implemented.
Documentation and availability#
The public website is uselumen.app, hosted on Cloudflare Pages. Links to a public repository and binary releases will be added when they're configured. Original source code and design use the MIT license; third-party artifacts keep their own terms.
This page was updated on 18 October 2026 from the current project documentation. Development moves quickly. If a screenshot of an earlier build differs from your app, trust the guide for your build and what your app shows.