0.1.14 is the first patch of the understanding beta, built from what the first 0.1.13 Mac told us on its first day.

Collection is honest about permissions. When macOS has not yet allowed the microphone, Luminars now records the screen anyway and says, on Home, in the menu bar and in Settings, exactly which permission it is waiting for, with the one button that opens the right place; the engine is no longer restarted every minute while it waits. The speech model that transcribes your audio is downloaded by the app itself from a fixed, verified source, resumes if the connection drops, and shows its progress beside the collection state, so transcription is ready sooner on a fresh install.

The personal analyst keeps working. A disagreement between your Mac's findings and their cloud copy no longer stops the analyst; it resolves the difference and carries on, and large maps publish in pages. The analyst's run archives now reach Luminars' learning store ahead of the capture backlog, which is what lets us learn from real work and improve the engine. Older runs are pruned on your Mac, the archiver no longer rescans everything every 45 seconds, and the sharing switch tells you when the cloud did not take a change and offers to try again. Drafts and Home show findings only where the analyst is the one writing them.

Under the hood: the recording bar can reach only its own nine commands; Windows paths now mint file resources and the Windows test build loads again; a connected agent of your own runs the AI from the moment you choose it; Usage shows minutes, not amounts, since runs are included in your plan; and the console's company pages refresh without blinking.

Luminars now builds an evolving understanding of your work and turns it into processes and improvement proposals you can review in Home and Drafts.

• Start from onboarding answers, company context and captured activity; inspect original evidence when needed.
• Keep findings, corrections and unfinished investigations when changing the AI runtime in Settings.
• Share supported work findings with your company automatically, with an opt-out and individual withdrawal controls.
• Follow company learning progress and assess opportunities using editable effort, review, maintenance and setup assumptions.
• Preserve original capture and investigation archives for the supervised beta, with visible upload status and operator restore tools.

An improvement proposal is a preparation step, not an automation that already runs. Estimated capacity value is separate from measured savings.

Home is new. One line at the top tells you where your day stands, Up next shows the one thing worth doing now, and every process you set up walks its own path: check it, connect what it needs, try it once with anything a connected app would send held for your approval, then turn it on. The morning brings a short brief on the day that closed, and Luminars tells you when it finds something worth your time, a few times a day at most and never during a call.

Luminars now reads your work as it happens instead of at the end of a session. Activities shows what you worked on, each activity backed by quotes from what Luminars saw, and a written diary of the day; you can correct an activity in your own words or give it a thumbs up or down, and your feedback helps us make it better. The day now ends at midnight everywhere. The inbox is now Drafts: what Luminars found, what you told it and what is almost ready to run, each as one card with its next step; approving a found process opens it with the assistant, to fill in what is missing.

The interview writes its questions as they come instead of all at once, some questions take several answers, and your processes come back with proper titles and the real icons of the apps they use. Context collection now runs from your first run, before you pick a plan, and if you close the checkout page by accident you can reopen it or go back to the plans. When the AI stops answering, Home offers Check again. Under the hood: the capture engine's speaker and voice-detection models now ship inside the app instead of being downloaded on first start, Home opens faster, recordings older than your retention window are really deleted, the menu bar icon shows when capture is paused or off, Open Luminars in our mails now opens the app, and pasting a connector key can no longer overwrite another one.

Video calls no longer weigh on your Mac. Luminars used to transcribe both sides of a call with its large model for the whole call, rebuilding its model every thirty seconds; now one transcription worker keeps its model loaded, skips a redundant pass, and transcribes the call in batches instead of live. Calls taken in the Comet browser are now recognised as calls, quiet speech is no longer dropped, and the capture engine runs at a lower priority than the apps you are working in. The app also ships its own audio and video tools, so capture never has to download them.

The interviewer now speaks its questions aloud (switch the speaker off at any time), and an interview you leave without saying anything closes cleanly. Runs can now read what Luminars remembers about your work and your process, and when a run does something differently from the process it tells you and proposes the change, which you can undo. When a run goes wrong, Luminars offers to diagnose it and proposes one fix. If you use Codex on a ChatGPT plan or Claude Code, runs and the assistant can work through your own agent. Hours and money saved now count only the runs that succeeded.

Luminars now finds the work you repeat by the shape of the work, not by its subject: a standing call with the same person, a pull request review, landing a coding round, a weekly update, a sweep of the inbox, whatever the project or client of the day. Each closed day is read as work units at that grain, the month's recurring kinds are listed once a week, and every unit of every day is labelled against them with the reason written down, so what Luminars found (and what it did not count, and why) can be read later. Your interview and your company's mapping now live on the plane: a browser interview from a mailed link, the org chart and roles in the console, the same engine in the app. Installations are now one per person, and everything the app observes about your work keeps reaching Luminars to improve the product (never the raw video or audio).

Luminars now keeps itself up to date: it downloads new versions on its own and installs them while your screen is locked, never during a recording, a run or the day's summary (Settings → General to turn it off). A one-click pause now lasts one hour, and Luminars tells you when collection has been paused for a while. The first interview with Dolly no longer wastes calls or your AI allowance. What your company knows now syncs and reaches the assistant. A new welcome and sign-in, a branded installer, and more of what the app observes is shared with Luminars to improve the product (never the raw video or audio).

Restores cloud sync for packaged installs so installation heartbeats, onboarding progress, diagnostic events, logs, and process activity reach support. Improves sync failure reporting and uploads logs every five minutes.

Everything below has been on main since 0.1.6 was cut on 18 September. No install has had any of it.

The audio your day is built on was being heard by the wrong model. Since the first launch in August, every Luminars install transcribed ambient audio with the smallest Whisper model and let it guess the language freely. Measured on a real day, 11% of what it heard came back in Korean, Japanese, Chinese, Cyrillic, Turkish or Arabic from one microphone in one room, and Italian speech came back as confident English nonsense. Nothing on the launch line had ever named a model or a language, so the engine chose one from the size of the Mac and kept it forever. This release names both: a large Whisper model and the six languages Luminars speaks. Your first launch after updating downloads that model once (about 830 MB, in the background); everything the day log, the suggestions and the search are built on gets better from there.

Your own agent can run the AI. Beside Luminars Cloud and your own provider key there is now a third way: a coding agent you already pay for. Codex signs in with your ChatGPT plan; Claude Code takes an Anthropic API key, or, if you say so explicitly, the Claude Code sign-in already on this Mac. The agent's own binary makes every model call, and it can reach Luminars' own tools while it works. Settings > AI, "Who runs the AI".

Settings has an AI room. The provider and your own keys left Account, where they were hard to find, for a room of their own. Account keeps sign-in, workspace and plan.

Luminars tells you when the AI side is down. Home shows one band for the AI channel beside the collection card, read off what actually happened rather than a guess. Every day close is now written down, a day that was owed one is paid back automatically (and on Catch up, and on a day's Retry), and each day carries its own word in the Activities picker with a Retry beside it. Days that were silently skipped while the channel was down come back on their own.

The run page, redrawn. A dark band for the head, the way back, Stop and the technical details in one nav row, the state beside the run's own clock, and the plan as a row of marks that names the step the run is on. Each started step is a card you open on that step's conversation, the composer sits inside the step the run is waiting in, and a run's outputs are cards you can act on. The page no longer jumps back to the top while a run streams.

A chat reads like a chat everywhere. The rail, the onboarding interview and the run page were three different drawings of the same thing; they are one now. The rail gained an empty state that offers three things to ask, Previous chats as a searchable list grouped by day, and a draft that survives switching threads. The interview says what the assistant is doing between questions. The run page folds a burst of tool activity into one group and draws a held step as a card asking for your approval.

Luminars can open a turn itself. A turn the product starts is marked as such, carries the evidence it is built on as chips into that day's Timeline, and closes with one press.

Notifications, rebuilt. The page, the badge and Home's band came from three derivations that disagreed, so the badge could never clear and a run parked on a human step was toasted and then shown nowhere. One source now feeds all three. The page is day groups of small cards, newest first, each with its own verb and Mark as read, and Mark everything as read in the head.

Processes and runs, after two walks. The trigger chip opens on every status. The list is one searchable, paginated table everywhere. Suggested carries a Confidence column with the basis behind each figure. Describe a process opens a chat that interviews you goal first and lands a draft. Settings > Skills imports every agent skill on this Mac in one click. On the runs list: the process name leads the row, one time column instead of two, a three-dots menu on every row with Stop for anything unfinished, and a Scheduled tab listing the next run of every process on a clock. The run clock no longer counts the time a run spends waiting on you.

Billing, and a recording you start by hand. While a grace period runs the app carries a bar saying so; when it is over the window is one billing screen until it is paid, and background collection and scheduled runs stop there. Nothing on your Mac is touched, and paying brings it all back. A recording you start by hand now runs even when background collection is off, whether you paused it or your workspace did; the engine comes up for that recording alone and goes back down when it ends.

Settings, shorter. About stopped being a room; its build facts and support address close General. Eight rooms, and the build facts show even if the rest of the read fails.

For workspaces. Team gets two hours of run time a seat, and the allowances were re-sized on what the work actually costs. The console's people are one tab now, and every form there opens as its own flow.

Everything below has been on main since 0.1.5 was cut on 5 September. No install has had any of it.

Luminars has a face. The new mark, a ring holding a shadow, is now the app icon, the tray icon and the first-run lockup.

A run can do far more on its own. A run now works in one persistent home with your own environment, and brings what it needs: Python rides inside the app beside the existing runtime, so a run reads PDFs, spreadsheets and Word documents with nothing installed. A step can use the browser (your own Chrome with a session you imported, or a browser built into the app) and the desktop itself. When a script or a recorded browser route has worked once, the next run replays it instead of asking the model again. A step that needs you is prepared by the run, handed to you with a notification, and checked when you are done. What a run installs is written down and can be reverted. No step waits for approval unless you mark it "ask first".

Work can start a process by itself. A watched folder or a mailbox can trigger a process: a file dropped in the folder, or an email that arrives, starts a run that carries it.

A skill folder becomes a process. Import an agent skill folder and Luminars turns it into a process, its files kept and its outputs previewed, asking only what a person actually knows.

Drafts start from the goal. A drafted recording now states the goal first and chooses how each step gets done, and the assistant's process panel asks that question before anything else.

Your assistant is yours. Choose Dolly, Ruby or Teo, each with a name, a colour and a living mark. The assistant holds more than one conversation, takes a file you hand it, listens when you speak (the microphone beside Send), knows what day it is, and answers sooner: it is already awake when you press Send, and a debrief turn takes two model calls instead of six. Under each reply it says how long it took and what it cost.

Suggestions listen to your firm. What your workspace says about its work counts as evidence, and a suggestion says "your firm said" when that is where it came from. Nothing is suggested until it has happened at least twice.

A daily briefing. Home and the Timeline open with a recap of your day, kept rather than rebuilt every time something new arrives.

Settings, rebuilt. Seven rooms behind the sidebar, with the Company page folded in. Every screen's wording was rewritten plain, Italian first, and the welcome screen has a language switch so the very first run can be read in Italian. The collapsed sidebar is an icon rail, loading states are skeletons, and an error shows a real page with a report you can copy.

Notifications arrive with the window closed. Every notification now comes from the part of the app that keeps running, and it tells you when context collection turns on or off.

For workspaces. A workspace has its own picture. An admin can turn context collection off for every install in the workspace from the console, and each app says why. Run time is an allowance per plan, Usage counts the AI by purpose, and when a budget stops something the app names which budget and for how long.

Fixes you may have met. A capture refusal that clears within a second no longer holds collection down for five minutes. The interview always has a way out. Signing out really leaves the app, and a session the service refuses signs you out instead of pretending.

Housekeeping. The app now sends its own log and the recording engine's log to Luminars every hour, so a problem can be looked at without asking you for a support file. The download is larger (about 300 MB) because Python and the desktop tools ride inside it. The recording engine is unchanged (Luminars Engine 0.4.46).

Everything below has been sitting on main since 0.1.4 was cut on 27 August. No install has had any of it.

The engine that records your screen is now ours. Luminars used to bundle the engine its vendor published; the vendor stopped publishing in August while a fix Luminars needed sat unreleased. Luminars now builds that engine itself from the source, signs it with its own certificate and ships it as Luminars Engine: the process a person sees running is luminars-engine. It carries the fix for a defect found on the owner's own machine, where clicks and keystrokes lost the app and window they happened in on five of twelve recorded days. Whether that defect is gone for good is decided by a full day of ordinary use on this build, not by this note.

Luminars speaks Italian. The app, the web console, onboarding, the permission dialogs macOS shows, the run transcripts and the exported workflows follow the language you choose, separately from the language your work is in. A real defect was found on the way: a bare language tag made one model answer in Finnish. It is fixed.

A new front door. Onboarding is one door: a short interview about how you work, then the app says it is ready before it asks for anything. Installs that predate the interview get a door into it.

Processes are easier to live with. A process can be archived and brought back. Any connection can be attached to a process, not only the ones its steps name. The "by hand" figure follows the evidence, and a recording you told the app about asks how long the work takes. The trigger speaks plainly, with a time-zone picker and the next run in words. Each process carries a picture. Tables are plain again, everywhere.

Luminars can now notice a job you repeat. Beside the three kinds of suggestion it already made, the miner recognises a job you do more than once and proposes it as a process. When a stated task exists, the suggestion stands under it in your own words.

The day log reads what you did, not only where you were. The day's record now carries cited actions with the thing they produced, drawn from the text you typed, the things you clicked and the windows you were in, rather than from window titles alone.

The AI keeps working on a busy hour. The pass that turns a recorded day into a readable log had a fallback model its provider had retired; it now falls back to a model that exists, and a second fallback rides on another provider. Hosted AI goes through one gateway with a model catalog that crosses providers. The daily allowances are doubled.

Intel Macs get a download. Luminars_0.1.5_x86_64.dmg is the first Intel build on the channel. It is built and signed by the same recipe and boots under Rosetta on the owner's Apple Silicon Mac; no Intel machine has run it yet. Treat it as a build to try, and say what happens.

Housekeeping you will not see. The engine's supervisor restarts it once, bounded, when its attribution goes blank. The release recipe no longer prints the update-signing key in the terminal. Windows is built and proven on a runner, unproven on a laptop, and is not on the channel.

Everything below has been sitting on main since 0.1.3 was cut. No install has had any of it.

A machine whose signed-in person changes keeps working. If someone else signed in on a Mac that had already been used, that machine quietly stopped syncing everything — usage, log slices, support files, consents. It said nothing while it happened. Fixed on both sides: the machine takes its own record over, and the server is the one that decides.

Your own OpenAI key works again. If you run Luminars on your own provider keys rather than the hosted AI, chat, drafting, running a process, the debrief interview and extraction were all failing — OpenAI renamed a field and the app still sent the old name. Everyone on hosted AI was already fixed on the server; this is the app's half.

Luminars is told when macOS turns a permission off. macOS takes screen recording back periodically, and until now the app just quietly recorded nothing. It now appears under "Needs you" with the way to turn it back on, and raises a notification. Two more things join it there: a connection about to stop working that cannot renew itself, and a draft whose questions you never got back to.

A connector's answer is fenced. What comes back from a connected tool is now clearly marked as data rather than instructions, and the AI is told what the markers mean.

You can say what happened. Settings → Your data → Report a problem now has a box for a sentence, sent with the support file. support@luminars.io is in the account menu, in Settings, and in the install guide. The support file itself is unchanged: still counts and states, still nothing you wrote or said.

"How much AI have I used" is in the app. Settings shows the workspace's month and your day against the plan.

Your work is backed up to Luminars while the switch is on. Everything the app produces — drafts, answers, conversations, run transcripts, recording bundles, day logs — now reaches Luminars' own copy under the "Help improve Luminars" switch in Settings → Your data. Turn it off and nothing goes.

Luminars now lives at [luminars.io](https://luminars.io). This release is the one that knows it.

• The app talks to luminars.io — sign-in, AI, sync, and the update feed. The old luminars-cloud.vercel.app address stays valid forever, and this build still falls back to it, so an update can never strand a machine on a name that moved.
• Your workspace's console is at [luminars.io/app](https://luminars.io/app). Every link you were ever sent still resolves.
• The site has a front door: what Luminars is, and a way to ask for access if you do not have a code yet.

Nothing else changed in the app. If you are on 0.1.1 or 0.1.2, Check for updates from the account menu offers this one.

## Luminars 0.1.2

Six small things a person meets in their first week, fixed before anyone meets them.

• Disk space. Screen and audio recordings are now kept for two weeks by default, then removed. The text Luminars already read from them stays, so your days remain searchable and your day logs stay complete. Choose 1 week to 3 months, or forever, in Settings → Context Collection.
• Open at login. The last first-run screen offers to open Luminars when you log in (on by default), so your days are collected without you remembering to open it. Change it any time in Settings.
• Your name. First run asks what Luminars should call you; Home greets you by it.
• The Timeline shows your clock. Times render in your own time zone, not UTC.
• The plain-draft screen — what you see when the AI's version of a recording does not hold up — has been reviewed and reads as designed.
• The menu bar's pause row says what it does: "Pause collection for 1 hour".

Apple Silicon only. macOS 13 or later.

## Luminars 0.1.1

The first published release, and the first one that arrives through the app's own update channel.

What is new since 0.1.0

• First run as designed — an invite code on the welcome screen, sign-in with a six-digit code (or the link from the email), your workspace, all four macOS permissions with real buttons, and an honest "you are set".
• Home — the landing surface, rebuilt; the assistant has a name (Dolly).
• Recording inside the app — a floating pill over your own tools, pause / mute / discard / stop, the drafting sheet with its stages, and Dolly on the first open question.
• Connectors need no setup — the app carries its own JavaScript runtime and its own copy of every connector server it ships; nothing is fetched from a package registry at run time. GitHub and Slack connect with one click through the hosted exchange (it stores nothing).
• Exports — a process leaves as an agent skill (a folder any harness can install) or an n8n workflow, each with a per-step capability report saying what the target can and cannot do with it.
• Support file — Settings → Your data writes a content-free support file and shows it in Finder; a second button sends it to Luminars. Nothing leaves your Mac unless you press it.
• Updates — "Check for updates" in the account menu works; updates are signed and install on your click.
• Fixed — the app inside the dmg now carries its notarization ticket, so a first launch works offline.

Apple Silicon only. macOS 13 or later.