Changelog

All notable changes to Lumina are listed here. The format follows Keep a Changelog and the project uses Semantic Versioning.

0.2.1 - 2026-11-01

Added

• Ask Lumina acts when you ask: creates, edits and deletes notes in Sources (deleting asks first), rewrites a meeting's summary, keeps attachments in Sources; every action has a card, most with Undo.
• Attachments: PDF, Office, text, audio and video in the chat, read or transcribed on your computer, optionally kept in Sources.

Fixed

• Opening after an update no longer needs lumina doctor: the old server is closed and waited for.
• Index again always indexes again and goes first.
• Folder file lists: the scrollbar no longer covers the delete buttons.

0.2.0 - 2026-10-31

Added

• Summary layouts (Settings → Summary): Standard, Executive brief, Class notes, Detailed minutes, Action plan, or Custom in your own words (blank = Standard).
• Web pages as sources: paste an address, read the page, then keep its content or a summary by your notes model. Cited with its address. MCP tool add_web_page.

Changed

• Folder file lists: name with Index again and Take out on one line, passages, size and state underneath, full width, no background.
• Chat: send and stop share one button.
• Imported recordings: the summary uses the whole width of the page.

0.1.8 - 2026-10-28

Changed

• Live assistance answers real questions: only complete questions on the meeting's subject get a card (no rhetorical "right?", half sentences or small talk); the answer comes from your sources, listing only the ones used, or from the web when they do not have it (only the question is searched; switch in Settings → Transcription). Models that think out loud no longer show their reasoning.
• On/off switch for live assistance next to its title in every meeting.
• Enhanced notes read like your notes: plain sentences that expand what you wrote, no "they talked about…", no bullets; a note never empties or shrinks on a later pass.
• Summary box: cleared after the new summary is written, and it grows line by line.
• Sidebar: Settings at the very bottom, under "Everything stays on this computer" and a thin rule.
• Notebook view: the typing hint sits under the notebook; the transcript bar keeps the page's edges.

Fixed

• Imported recordings say "Transcribing" at once, in the meeting and the list.
• Transcribe again works and cannot start a second run over one in progress.

0.1.7 - 2026-10-28

Added

• Summary the way you want it: a box under the summary to say how you want it (an email, only the decisions, another language…) with Regenerate summary, plus Expand, Shorten, Simplify, More formal and Action items first. For finished meetings and imported recordings.
• Connected apps: save a meeting to brain (or every meeting, automatically) and search brain from Ask Lumina; send a meeting to an OpenCue agent and read its reply; see and open AI WorkStudio. Settings → Connected apps.
• MCP tools rewrite_summary and save_meeting_to_brain.

0.1.6 - 2026-10-25

Changed

• macOS asks for permissions once, in the walkthrough. The setup's last step (and Settings → Permissions) has an Allow button for the microphone and one for the computer's audio. Starting a meeting never opens a permission window; a missing permission shows a notice with a link to Settings.
• The permissions survive updates: they belong to Lumina's own audio helper (data/bin/LuminaAudio.app), which updates do not replace, instead of the desktop app, which every update re-signed.
• Computer audio without screen recording on macOS 14.2+: a Core Audio process tap needs only System Audio Recording.
• Sidebar: the Recording, indexing and Update cards sit right above “Everything stays on this computer”, which now fits on one line.

Fixed

• Updates failing because of uv.lock: the file is no longer part of the code and the updater discards a local copy. Coming from 0.1.5, if the update fails, delete ~/.lumina/uv.lock once.
• Tablet layout: two-column settings line up (Settings → Knowledge), the notes-view cards align, the Settings index hides below 1100 px and the meeting header keeps its buttons on the date row.

0.1.5 - 2026-10-24

Added

• Regenerate summary on every finished meeting.
• Imported recordings have their own page: player, summary and transcript, with Transcribe again; no notes columns.
• Stop in the chat above Send (what was written is kept), and Answer again / Edit question under your last question.
• Right-click on chats and meetings: Rename and Delete.

Changed

• Enhanced notes are only your notes, and automatic. No Enhance button and no "Lumina ·" notes of its own.
• Faster enhancing: the model reuses the part of the transcript it already read, only your cues are answered, and it does not run when you wrote nothing.
• Ending a meeting does not enhance; it cancels any enhancing and writes the summary. Enhanced notes and summary are separate.
• No scrollbar in the chat list.

Fixed

• macOS permission windows (best effort): the app is signed ad hoc as a whole so macOS remembers the permissions; lumina doctor shows their state.

0.1.4 - 2026-10-24

Added

• Update from inside the app. Lumina looks for a newer version when it starts and every 6 hours and shows an Update to x.y.z button under the menu. Settings → About has Check now and a switch for the automatic check. Updating runs lumina update for you, closes Lumina and opens the new version.
• Meetings in Sources. Finished meetings, with their summary, are listed in Sources so everything Lumina can search is in one place.
• Free memory when idle (Settings → General, on by default). The speech, embedding and built-in notes models are unloaded a few minutes after you stop using them and load again by themselves.

Changed

• New, simpler logo: an L and a dot of light, in the app, the website, the app icon and the share image.
• One program asks for the Mac permissions. The microphone and the computer's audio go through the same signed helper, so each permission is asked once. The helper is rebuilt only when its source really changed and is signed with a fixed name; if it cannot give audio, Lumina falls back to the previous microphone path.
• The engine follows the app. If the app closes or crashes its server closes with it; if the python process is closed by hand, the app starts it again.

Fixed

• "The model did not return valid JSON" in Enhanced notes and the summary. Answers with extra text, a trailing comma or cut short are repaired, and otherwise Lumina asks again with more room and without the JSON grammar.
• Sources no longer breaks with long names of files, folders, paths or indexing messages.

0.1.3 - 2026-10-24

Added

• Indexing progress bar. Sources shows what Lumina is reading or indexing, files done out of total and a rough time left; a thin bar under the sidebar shows it on every page and folders show their own.
• Folders open like a dropdown. Click the arrow on a folder to see every file inside with its status, passages and size.
• Settings → Delete data. Delete all chats, all sources, all meetings or everything, after a confirmation. Passages go with them; settings and downloaded models stay.
• Live transcript in the notch (desktop app). While you record, a black island grows out of the notch (or the top edge of the screen) with the last three lines of the transcript. Click it to return to the meeting. Turn it on or off in Settings → Transcription.

Changed

• Much faster indexing. Files are read while the model embeds the previous one, PDFs are read with PDFium (about 6 times faster than pypdf on a plain file), passages are batched by length, repeated passages are embedded once, unchanged files are skipped and the model loads while the first file is still being read.
• The website uses the app's look: near-white desk, hairlines, charcoal buttons, soft blue, neutral dark theme, new logo.

Fixed

• Deleting a file or folder from Sources now deletes its passages and the counter updates at once. A file being indexed while you removed it could write its passages afterwards, dropped-file copies stayed in data/uploads and the page never received indexing events. Passages without a source are removed at every start.
• A damaged file is reported on its own row instead of stopping the queue.

0.1.2 - 2026-10-23

Fixed

• lumina update and the install command now really update. Copies installed without git history (the first installs) were never refreshed and kept reporting 0.1.0. The installer now downloads the current code once, keeps your data/ folder and turns the folder into a normal git checkout. Same fix in install.ps1.
• UI positioning. The meeting title has its own line instead of hiding under the buttons; the sidebar collapses to icons below 1000 px; the meetings table no longer overflows on narrow windows; stray rounded corners, the tag field and the warning banners were cleaned up.

Added

• Stop the local model from the app. Settings → Models has a Running now panel with the notes and chat model, the speech model and the search model, each with a Stop button that frees its memory. They start again by themselves when needed. Settings → Notes model also has a Stop button.

Changed

• A lighter, more minimal look: white sidebar, near-white desk, hairline borders, softer shadows, a soft blue used only for links, focus and selection, and charcoal primary buttons. The dark theme uses neutral greys.

0.1.1 - 2026-10-23

Added

• Three views for meeting notes, chosen in Settings → General with a small drawing of each: Side by side (the new default: your shorthand notes on the left, the live enhanced notes on the right, like Google AI Edge Foresight), Cornell rows (the 0.1.0 layout) and One at a time (a wide page with a switch between your notes and the enhanced ones).
• EmbeddingGemma 2 as a speech option (experimental). EmbeddingGemma 2 is an embedding model and does not transcribe by itself, but its audio encoder plus a small community reader (mlboydaisuke/EmbeddingGemma-2-Transcribe-LiteRT, Apache-2.0) does: about 4 % word error rate on English read speech, on the CPU through ai-edge-litert. English only, no punctuation, each phrase appears when the speaker pauses, about 1 GB. Never chosen by "Automatic".
• Models grouped by what they do in Settings → Models: Chat and notes, Transcription, Embeddings and Read aloud.
• A real harness for the chat: a structured system prompt (grounding, citations, language, style, prompt-injection rules), follow-up questions rewritten into standalone searches, retrieval without duplicates and with a budget per source, a digest of recent meetings, the full notes of the meeting a chat is about, tagged sources that cannot close their own tag, and a clean-up pass for citations.

Changed

• New chat design: minimal and quiet, with answers as plain text, sources as small chips and a floating composer.
• The default notes view is now Side by side; choose Cornell rows in Settings → General to keep the old one.
• ai-edge-litert is a new dependency (Windows, Linux and Apple Silicon Macs; Intel Macs keep every other model).

0.1.0 - 2026-10-22

First release.

Meetings

• Live transcription of the microphone ("You") and the computer's audio ("Others": Zoom, Meet, Teams…), on the CPU. macOS captures system audio with ScreenCaptureKit; Windows with WASAPI loopback; Linux with the PulseAudio/PipeWire monitor.
• Speech models: Parakeet TDT 0.6B v3 (25 European languages, detects the language), Canary 1B v2 (same 25 languages, follows the language you choose) and Whisper tiny to large-v3-turbo (99 languages). Silero VAD cuts the audio into utterances; words show up as they are spoken.
• Transcription language selector with Automatic plus every supported language. With "Automatic" engine, Lumina picks Parakeet for automatic detection, Canary for a fixed European language and Whisper for the rest.
• Cornell-style notes: your shorthand in the cue column, Enhanced notes aligned row by row, extra points Lumina noticed, and the summary (overview, decisions, action items, open questions, topics) at the foot of the page.
• Live Enhance during the meeting (automatic or Enhance now) and a final pass when the meeting ends; automatic meeting titles.
• Live assistance: questions asked in the meeting that your sources can answer get a short, cited answer card.
• Pause, resume, end; import an audio or video recording; full transcript with search; tags; filters by text, tag and date.
• Export to Markdown, plain text, SubRip subtitles (.srt) and JSON.

Knowledge

• EmbeddingGemma 2 (8-bit ONNX, 768 dimensions, truncatable to 512/256/128) downloaded from the app, running in process.
• ChromaDB collections for documents, meetings and notes; meetings are indexed when they end.
• Sources: files, folders watched for changes (watchdog), uploads and notes. PDF, DOCX, PPTX, XLSX, Markdown, text, HTML, CSV, JSON and code. Tags per source.
• Chat over everything with numbered citations, history, an optional Web search toggle (DuckDuckGo or SearXNG) and "Ask about this meeting".

Models

• Notes model: Gemma 4 E2B, E4B or 12B (4-bit GGUF) run by a pinned llama.cpp server (b11823), or Ollama, LM Studio or any OpenAI-compatible server. Ollama and LM Studio are detected automatically.
• Every model shows its size, license and memory needs before downloading; downloads resume and can be cancelled.
• Optional read-aloud with Kokoro 82M or the system voice.

Agents

• MCP server with search_knowledge, list_sources, get_document, add_note, query_meetings and get_meeting, over stdio (lumina mcp) and HTTP (/mcp, bearer token). Ready-made snippets for Claude Desktop, Claude Code and Cursor in Settings → Agents.

App and install

• One-line install: curl -fsSL https://lautaro005.github.io/lumina/install.sh | sh (macOS, Linux) and irm https://lautaro005.github.io/lumina/install.ps1 | iex (Windows). lumina opens the browser, lumina app the desktop app.
• Desktop app (Tauri 2) with a tray icon; it starts and stops the local server.
• First-run setup: interface language, meeting language, notes model and downloads.
• English and Spanish interface; light and dark themes.
• Website with documentation, privacy policy, security policy, changelog, llms.txt and sitemap.

Known limits

• Installers are not code-signed or notarized yet: macOS may ask you to confirm the first launch, Windows SmartScreen may warn.
• No speaker diarization beyond You / Others.
• The Linux desktop build is x86-64 only; other Linux machines use lumina in the browser.