Voice profiles hold up better. Merging two people into one profile no longer degrades recognition of either. Re-importing a recording no longer adds duplicate samples. If your profile stops recognizing you after a mic or room change, "Is this you?" comes back even when a colleague was recognized; confirming rebuilds the profile from that recording, with one Undo that reverses the link, the samples, and the rebuild. Every identity edit now follows that rule: one gesture, one ⌘Z.

Transcription drops fewer words. A rebuilt Parakeet encoder replaces the stock build, which corrupted tokens under certain conditions. Existing installs fetch it in the background while idle and switch on the next launch. Short replies ("yeah", "mm-hm") are credited to the right person more often instead of parked as Unknown, and speech at the start and end of a recording is labelled more reliably.

Other improvements:

• Flagged Moments quote the full paragraph you marked, in the app and in exported notes
• Failed links, profile rebuilds, and cross-meeting link offers now show in the overlay and summary window, not only the history window
• ⌘Z right after a rename waits for the summary refresh and applies, instead of being ignored
• Speaker Profiles in Settings updates live
• Custom endpoints say why a model list didn't load, with a link to the Network Log
• The dictation waveform stops redrawing while flat
• Elastic scroll bounce on every column
• Fresh installs no longer depend on a model file upstream stopped publishing
• Updater moved to Sparkle 2.9.6

Bug fixes:

• Renaming a speaker no longer flashes the old name
• Switching summary views no longer revives a retired speaker name
• Long bookmark excerpts no longer paint over the summary
• "Is this you?" no longer proposes a speaker you named as someone else
• A failed recording start no longer blocks the next one
• The permissions alert says which pane opens when both are missing

Improvements & Fixes:

• Near-simultaneous speech across the mic and system channels is re-sequenced so a speaker's turn stays in one piece in the live transcript
• Clear speech after a pause survives end-of-recording trimming, without phantom words coming back with it
• Live segment merging recognizes quoted dialogue, parentheticals, and abbreviations, so a turn is no longer cut mid-sentence
• Timestamps no longer run backwards across merged live transcript chunks

Speech decoding across live listening windows is substantially more stable. Sliding-window boundaries now snap to natural voice activity valleys, cutting during silence pauses rather than slicing mid-word. Prefix-seeded TDT decoding primes the language model across window seams, fixing a long-standing failure mode where multi-token numbers (like "582,000" or "$2M") and compound terms broke across seams. Live transcription churn is visibly reduced: a new two-tier display distinguishes committed text from provisional words, letting the live transcript settle faster while preserving 90.7% sentence punctuation precision and 95.6% boundary recall at speaker handoffs — tracked on the public live transcription benchmark.

Meeting summary generation is roughly 3x faster (dropping post-meeting enrichment latency from ~12s to ~4s in benchmarks). Because on-device speech recognition and formatting are now accurate enough on their own, the post-meeting pipeline no longer requires a heavy LLM pass to re-read and echo back the full transcript. Instead, summarization, speaker attribution, and action items now execute in a single streamlined prompt pass. Eliminating the full-transcript echo requirement cuts input tokens by over 60% and output tokens by over 70%, allowing post-meeting processing to run far more reliably on smaller models and local endpoints (via Ollama or LM Studio) without running into timeouts, context bottlenecks, or output truncation.

Flag important moments with one tap. A new bookmark button and shortcut let you flag key points during a live meeting or mark any line in transcript history. Bookmarked moments are highlighted in the transcript, guaranteed to be covered in the AI meeting summary, and automatically included across Markdown vault notes, webhooks, and JSON exports.

Other improvements:

• Multi-format transcript export: save any meeting transcript as Markdown, VTT, SRT, or structured JSON.
• Audio file import directly from web URLs (HTTP/HTTPS).
• Real-time audio capture health indicators: an amber badge warns when audio capture falls behind, and a red banner alerts if audio stream frames are dropped.
• Streamlined LocalVQE audio processing thread pool, cutting streaming audio CPU overhead by 50%.
• Custom LLM endpoints now stream progressive transcript reveals during meeting enrichment.

Bug fixes:

• Fixed seam-stitching whole-channel fallback to splice missing audio spans rather than replacing the live stream, eliminating intermittent blank gaps after pause recovery.
• Live toolbar controls maintain a fixed width, preventing unexpected UI jumping during meetings.
• Stutter-aware prefix priming prevents the prediction network from over-penalizing natural conversational stammers.
• Fixed a race condition where rapidly toggling bookmarks on the same segment could duplicate or drop entries.
• Resolved speaker sample foreign-key parent resolution during database migrations.
• Corrected live transcript scroll locking so the view reliably tracks the tail of new speech until manually scrolled back.

Improvements:

• Quiet microphone speech is more likely to survive heavy system audio. The gate that keeps the other side of a call out of your transcript reads a word's loudest instant rather than a flat −50 dB average.
• Words are judged whole rather than piece by piece, so a word spanning two moments of audio arrives intact instead of clipped short or run onto the one before it.
• When the decoder goes blank over speech that is plainly there, that stretch is re-read with its text context cleared. Most of what comes back is one speaker talking over another.
• The live meeting Q&A moved into the same scroll as the briefing and running summary, laid out like the post-meeting window. Live questions and answers are saved with the meeting now.
• Every setting sits on the pane whose feature it affects. The AI pane is now "AI & Data", beside the Network Log in a Privacy group.
• Controls the routing has already decided now lock rather than offering a choice that changes nothing.
• MimicScribe no longer sends notifications. Automatic-recording rows are a switch per app, and a failed backup warns on the hub instead. The one-tap Discard on a self-started recording goes with them.
• Overlay panels have soft scroll edges and a native blur under their bottom controls.
• The live transcript follows new speech and stops following once you scroll back into history.
• The response overlay's default font size is 17pt. An existing choice is unchanged.

Bug fixes:

• A question typed into a meeting Q&A input could vanish, most often when asked from the transcript history window.
• Talking points and action items could fail for an entire meeting with no sign on screen, including when no API key was set. Failures now surface, and a panel that already has content is left alone.
• "Transcript only" disappeared behind a loading skeleton for the whole finalize, then reappeared at the end.
• "Keep transcript only" on a failed summary left the same error headline and the same button in place.
• The audio level overlay tracked the frontmost window rather than the field you were typing into. (#14)
• The Network Log painted every content row orange on the default cloud setup, and described a webhook as carrying metadata only.
• Clearing the meeting shortcut during onboarding left Next disabled with no way forward.
• Mic and system icons in the live transcript sat below the line of the text beside them.
• Section headers and dividers in the live assistant column sat slightly out of line with the transcript beside them.

Saved voice profiles recognize returning speakers more reliably. On the AMI corpus, 92.0% of returning speakers are matched to their own profile across 779 trials, with zero wrong identifications (95% upper bound 0.5%). The same production decision rules now run against a second corpus, ICSI, also with zero wrong names. The rule throughout is refusal over guess: when two profiles score too close to call, the app declines and marks the speaker instead of writing someone else's name into a durable transcript. A profile also holds separate voiceprints for genuinely different conditions, so AirPods and a laptop microphone no longer average into one blurry voice, and each voiceprint has to clear its own confidence bar before it can claim a match. Enrolling a profile now takes 20 seconds of speech instead of 30. Full results and method.

Recognition now works against two bars rather than one. Above the higher bar the app names the speaker itself. Below it, instead of guessing or going quiet, it offers a one-click "Is this X?" suggestion, and those suggestions are right 90% of the time. Naming someone also reaches backward: enrolling a voice sweeps your earlier meetings for that person, applies the name, reports what it did, and undoes the whole sweep in one step. Cross-meeting matches that fall short of claiming a name on their own arrive as a list you accept one at a time.

Automatic recording, in beta. MimicScribe can start and stop a recording when it detects a call, and stays off until you enable it. Teams, Slack, and Zoom are detected with no Accessibility grant at all. Browser calls are detected by the site in the tab, from a list you control, in Chrome, Chromium browsers, and Safari. Firefox reports itself unsupported rather than guessing. False starts and missed calls are the thing to report.

Meetings can be summarized by Apple Intelligence, on the device, with no network. Long meetings work. This is an early experiment rather than the recommended setup: the on-device model finds fewer action items than the cloud one does, so summaries written this way are labeled and carry that caveat.

Other improvements:

• Transcription window boundaries are placed at a quiet point in the audio rather than on a fixed clock, so fewer words are cut in half at a boundary in the first place. The rest of the word-recovery layer was retuned around it.
• Speech the transcriber skipped over is re-read at a slower rate, which recovers words a normal second pass cannot. This runs during live meetings as well as on imported files.
• Filter meetings by a saved speaker profile and search inside that filter. Your own turns are found before they are linked to a profile.
• Voice profiles are matched on imported audio, not just live recordings.
• Filler words are removed from the saved transcript, with the surrounding capitalization and punctuation repaired.
• Every summary records which engine wrote it.
• An offline meeting shows where it stands in your notes folder.
• Shortcuts and Siri actions have a master switch.
• The recordings sidebar no longer reads every transcript to draw itself, and launch, merge, export, search, and the meeting overlay do less work on the main thread.
• The transcript stays legible while a meeting is being summarized.

Bug fixes:

• A recording the app decided was silent could be deleted along with its audio. Two separate paths did this, at stop time and after post-stop analysis; both now preserve the audio for recovery.
• An on-device summary that stalled could leave a meeting stuck on "processing" for the rest of the session, and a run that gave up partway presented its partial result as finished. Both now end in a failure state with a retry.
• The microphone was briefly opened when the screen woke.
• A word landing on the boundary between two transcription windows could be committed twice ("Adamsams").
• Renaming a meeting destroyed the search index for its action items, and could leave the summary naming a speaker who no longer existed. It now offers a refresh and retags action item owners.
• "Undo automatic enrollment" left the profile it had created in place.
• Changes made by the automatic enrollment sweep no longer claim in the edit history that you made them.
• Two views reporting on your whole history were reading only the 50 most recent meetings.
• A Safari call could never end on its own, and automatic-recording notifications named the audio helper process rather than the app.
• Usage reporting followed the AI provider you picked rather than the toggle you set.
• Fixed a crash caused by a UI update off the main thread.

⌘Z undoes your last edit to a meeting and ⇧⌘Z redoes it, from any window that shows the meeting.

Click a speaker's name in the transcript to rename it in place. The rename reaches everywhere the old name was, and profile edits (delete, rename, merge, attach) are now reversible.

Other improvements:

• Select several speaker profiles at once in Settings and merge or delete them together.
• Speaker colors belong to the person rather than the capture channel, and the palette widened to 14 hues that hold their range in dark mode.
• The transcript marks a speaker who can be saved with a plus, and one already saved with a checkmark.
• Earlier meetings holding the same voice are offered as links when you save or attach a profile.
• Gaps where people talked over each other now fill in during the meeting instead of after you press stop.

Bug fixes:

• Renaming a speaker onto a name that already exists merges the two instead of hanging the correction.
• Undoing a merge restores the profile link and keeps the surviving speaker's name.
• Undoing one edit no longer rewinds past the edits you made before it.
• "Add to an existing profile" says out loud when it can't proceed, instead of doing nothing.
• An edit aimed at a profile that no longer exists is refused rather than reported as done.
• Stopping a long recording no longer discards audio the transcriber hadn't reached yet.
• VoiceOver no longer reads the "Unknown" label as if it were a speaker's name.
• The duplicate inspector toggle can no longer re-show itself.

Speaker labels while you record. The live transcript now separates speakers as the meeting runs — a numbered, color-coded badge per row — where before it showed what was said but not who said it. Labels are held back until the evidence settles, so you see a plain channel label instead of a guess you would have to correct later.

Voice profiles that hold up across meetings. A saved profile is now matched to a speaker only after the transcript is finished, so it can no longer attach to a stray fragment of someone else and carry that wrong name — and that wrong voice — into every meeting afterward. Two speakers who turn out to be one person can be merged from the transcript, and any merge can be undone.

Fewer dropped words. Words lost when several people talk at once, which the first transcription pass skips over entirely, are now recovered by re-reading just those gaps. Stray periods no longer appear mid-sentence where two windows meet, and long recordings stop losing audio when transcription falls behind live capture.

Other improvements:

• The enrolled-profile badge is now a button: "This isn't me", re-point to a different profile, or rename — the undos used to be buried in a per-turn menu.
• "Save voice profile" is offered after three good stretches of speech instead of five, closing a gap where a speaker with a minute of clean voice was never eligible.
• The outcome of a speaker action now appears in the recordings window where you did it rather than the menu-bar panel: a quiet receipt when it works, and when it doesn't, a banner that stays until you dismiss it and offers the fix as a button.
• Profile popovers and the profile picker: bigger text, readable captions, and action links with a real click target.
• Speakers eligible to be saved are marked in the transcript, and their badge names and saves that voice without leaving the meeting.
• Lines under two seconds never assert a speaker, and a speaker appearing for the first time keeps a plain channel label until the stable pass confirms it.
• A live speaker that turns out to be covering two different voices now splits into two.
• A caret marks speech that has been captured but not yet transcribed, so a pause no longer reads as a stalled recording.
• Saved profiles record which meeting enrolled them.
• The meeting header shows one date — when the meeting was recorded, not when the file was imported.
• Integrations settings regrouped, with MCP ahead of the rest and tightened descriptions.
• Post-meeting processing takes roughly a tenth longer, which is what the gap-recovery pass costs; recordings with fewer gaps to re-read pay less.

Bug fixes:

• Adding a speaker to an existing voice profile could silently undo the link it had just made; picking your own profile did nothing at all.
• Speakers the AI renamed became invisible to profile matching and to "Save voice profile", however long they had spoken.
• A period inside a word ("education. mp3") no longer splits the sentence in two.
• A locked keychain no longer strands the app on "Unable to connect to AI service" for the rest of the session — credentials are recovered mid-session.
• Databases left in a partially-migrated state are repaired on launch.
• A speaker with only a little speech is no longer absorbed into whoever spoke next.
• Fixed a crash when live transcript rows rendered outside the overlay panel.
• The History window no longer spins when a deferred meeting is parked.
• Live speaker colors mean identity rather than capture channel, and speaker numbers stay contiguous instead of skipping.
• The same Logseq graph no longer appears twice in quick-setup.
• The duplicate inspector toggle is hidden.

Write meeting notes straight into Obsidian or Logseq. Point MimicScribe at a vault folder and every finished meeting is written as a Markdown note — correct flavor per destination — with summary, action items, and participants. The app stamps each note it writes and stops overwriting once you edit it, so your copy is never clobbered. A one-click backfill exports existing meetings.

Per-meeting edit history, local snapshots, and encrypted backups. Every meeting keeps a change ledger: renames, merges, corrections, and summary edits are recorded, so you can undo, redo, or seek back to a lossless original of the meeting as first saved. Your database is also snapshotted automatically on-device, and you can export a fully encrypted backup — keyed to a recovery key only you hold — to restore on another Mac. (#50)

Spotlight, Shortcuts, and Siri. Meetings — full transcript included — are now indexed by Spotlight, so a query for something that was said jumps to the meeting it was said in. App Intents expose search, open, summary, and action-items actions to Shortcuts and Siri, so meeting data drops into your own automations. Indexing is on by default; toggle it in Settings.

Other improvements:

• Dictation and Transform moved to the new version of Gemini Flash-Lite to fix product-naming issues, where the model would replace current product and version names with older ones from its training data.
• Outbound webhook: POST summary, action items, tags, and participants (never the transcript) to your own endpoint on completion.
• A Models section in Settings lists downloaded speech/speaker models with size and location, plus a safe re-download if one is corrupted.
• Integrations settings regrouped with an explicit On/Off state per integration.

Bug fixes:

• Renaming a speaker onto another saved profile in the same meeting now merges the two profiles — keeping the target's voice model — instead of erroring that the name is already in use. This is the fix for correcting a person the diarizer split into two.
• No more brief launch hang from reading saved keys on the main thread.
• Opening a meeting from a Spotlight result reliably navigates to it from a cold start.
• Undoing a speaker reassignment counts the turns changed, not the underlying segments.
• Dictation preserves embedded asides and self-reminders instead of dropping them.
• Edit-history panel fixes: opens correctly in the History window, no stale state after reassigning a speaker, corrected summary-banner spacing.

Recover interrupted meetings. If the app crashes or a recording is cut short, your audio is no longer lost. The interrupted recording is kept and appears in the sidebar with a one-click "Transcribe Now" that runs the full transcription and speaker separation after the fact — and crash-damaged recordings that used to refuse to import ("No audio track found") are now repaired automatically. (#49)

Cleaner speaker labels, with controls when you need them. This release removes more of the phantom "extra" speakers that used to appear from played-back video and quiet cross-talk. Two new controls help when the automatic pass gets it wrong: turn on "My mic is always me" so your own channel is never split into several people, and right-click a speaker in the transcript to reassign all of their turns to someone else at once.

Better echo cancellation on speakers. On speakers instead of headphones, your mic re-records whatever is playing — the other side's voice, a video, a podcast — and without cancellation that echo lands on your mic track as a phantom duplicate speaker. MimicScribe's on-device neural canceller subtracts the system audio back out of the mic in real time. This release upgrades that model (cleaner cancellation at ~11% lower CPU) and closes a failure mode where a dropout in the system-audio tap would walk the echo past the canceller's alignment window and silently kill cancellation for the rest of the call — it now detects that drift and re-locks.

MCP can edit transcripts now. The MCP server used to be read-only — search, fetch, summarize. The new correct_meeting tool is its first write channel: an agent can rename, merge, or remove a speaker on a saved meeting, validated against the live roster so a bad suggestion degrades to a no-op instead of a wrong edit. Transcript surfaces also tag each turn's capture channel — mic, remote, or mixed.

Other improvements:

• Combine two recordings into one — if a call got split across separate recordings, merge them from the sidebar into a single meeting, interleaved in order, with notes preserved. You can also rename a meeting inline.
• Export a meeting's summary and action items, not just the transcript.
• Search results show which action items matched, update live as you type, and are keyboard-navigable.
• Dictation cleans up offline: repeated stutters are de-duplicated and pauses become paragraph breaks.
• The bring-your-own-AI custom endpoint is free during the launch promo, and MCP access is included.
• Resend your own license key from the website if you lose it.

Bug fixes:

• AirPods (and similar processed mics) no longer misattribute the first words of a turn. Their onboard audio processing reshapes the start of each turn, which used to push those opening words onto the wrong speaker or "Unknown."
• Speaker and meeting renames no longer revert when you reopen a meeting.
• Dictation no longer pastes a hallucinated phrase after a trailing silence.
• Fewer dropped words on quiet or dial-in audio. Faint and telephone-band speech sometimes slipped past the voice-activity gate — dropping whole stretches of words — or fell in the gap between transcription windows. The app now recovers both.
• Transcripts no longer drop a word where the speaker changes mid-sentence.
• A transient keychain hiccup can no longer temporarily downgrade a paid license.
• Losing microphone permission mid-meeting is now surfaced instead of silently recording nothing.
• Transform mode requires an AI provider — the confusing "Off" state is gone.

Long meetings get more complete summaries. The assistant builds the post-meeting summary from the running summary points it collects during the call, so a decision or number mentioned once in a 90-minute meeting is much less likely to drop out of the write-up. Transcripts and summaries also stream in as the model responds instead of landing all at once — the attributed transcript starts filling in a couple of seconds after you stop.

Search is rebuilt. The old version often missed what you were looking for; the new one spans your whole meeting history — type a topic or a person's name in the sidebar and get matches from every meeting at once, with an AI answer drawn from them. Speaker labels are searchable too, so "Remote 3" or a named participant turns up their own turns and the places they're mentioned. If you keep meetings offline by default, sending a search to the AI asks first.

Imported audio files now have their own place. File imports live in a dedicated sidebar category alongside your meetings, and honor offline mode the same way a live recording does. You can also flip a meeting offline mid-recording and back.

Other improvements:

• Edit a meeting's prep mid-recording without losing the live meeting or its transcript.
• Speaker labels number compactly — a two-person call reads "Remote 1 / Remote 2," and a lone remote is just "Remote."
• The Network Log now shows whether each request carried meeting content or only metadata, and marks whether your AI endpoint is on-device or remote.
• Action items drop the due-date line when no deadline was actually mentioned.
• Sidebar rows gain a right-click menu: Delete, Export, Reveal Recording, Copy.

Bug fixes:

• Long meetings no longer trip the "AI features paused" limit for Unlimited users.
• Fixed a beach-ball stall that could freeze the app during a live meeting.
• Hardened against a rare audio-engine crash when switching input devices.
• Dictation no longer drops the final words when you stop during a quiet tail.
• The post-meeting correction reply no longer reads a successful edit as a failure.
• The Assistant button now opens on meetings started without the assistant visible.
• Imported meetings are dated when you import them, not by the file's creation date.
• Rapid back-and-forth Q&A no longer shows up out of order in the saved transcript.

This release is about live transcription quality at the seams. Streaming transcription listens in overlapping windows, and where those windows are stitched together is exactly where words used to garble or vanish. The stitching logic was rebuilt around a strict never-drop rule, plus a new repair pass that re-decodes a seam when the committed text looks wrong — and the result is measured in the open: the live transcription benchmark is now public, counting every window-overlap disagreement across 8 AMI meetings — 736 stitch points, zero losing 3+ meaningful words, every one- and two-word loss counted — with real example stitches shown next to the human reference transcript so you can judge each one. Sentence-punctuation accuracy against human-punctuated Earnings-21 references is included.

You can now find meetings by who was in them. Search understands people — "meetings with Sarah about pricing" matches named speakers (including renamed and merged speaker profiles), requires both the person and the topic, and each result shows badges explaining where it matched: person, transcript, summary, action items, or topics.

Other improvements:

• Bring-your-own enterprise AI endpoints got another round of hardening: custom auth headers for API gateways, sovereign Azure clouds, a Mistral preset, and a Verify step that self-heals common configuration mistakes.
• Offline transcripts drop spurious mid-sentence periods ("and they pretty. soon" → "pretty soon") — validated against human-punctuated references.
• Lower memory use while recording: audio buffers are pooled and rolling tails use fixed-size rings.
• Minimizing the meeting HUD now pauses live-assistant work — no wasted compute or metered briefings while you're not looking.

Bug fixes:

• "Generate Summary Now" shows real staged progress with an accurate elapsed timer, and no longer shows a stale search prompt in history.
• Fixed a hang when stopping a recording after the live transcript had fallen behind.
• Short meetings no longer write an unnecessary second copy of their audio to disk during processing.
• A sentence fragment trailing its speaker by a beat ("…have to be.") is reattached instead of being dropped from offline transcripts.
• Recovered automatically from a database state that could prevent the app from launching after an externally-edited database.

This release sharpens what ends up in your transcript. The transcription model sometimes inserts names it "remembers" from podcast training data into unrelated audio — a stray "Aaron Powell" appearing as if someone said it; those are now detected and removed. Short bursts of speech that a decoding glitch used to drop mid-sentence are recovered from the live preview, quiet trailing words are kept more often, and numbers and spacing read better offline: "one billion" becomes "1 billion", "D V D" becomes "DVD", and a split "20 16" rejoins into "2016". Punctuation at transcription-window seams is cleaner too.

The live transcript now shows the whole meeting instead of just the last few minutes, and you can search it while the meeting is still running. It follows the newest text as it arrives and stays put when you scroll back to re-read. Several remaining rough edges are gone — it no longer freezes after a burst of background processing, doubles a phrase across refreshes, or briefly shows a hallucinated line during a pause.

Other improvements:

• Custom AI endpoints now measure the model's usable context window during setup and fail loudly instead of silently truncating a long meeting. A meeting summarized without the speaker-attribution pass shows a clear note with a one-tap "Refine now."
• Attributed transcripts requested over MCP return an on-device transcript even when no AI provider is configured, rather than failing outright.

Bug fixes:

• Fixed a rare crash that could quit the app during dictation (Insert and Transform modes). (#48)
• A recording with no speech is now discarded cleanly in the history window instead of leaving an empty, "completed" meeting that reads "No speech detected."
• Removed a duplicate transcript-inspector toggle button, and aligned the follow-up question box with the transcript search field.

This release moves live meetings into the main window. A meeting now opens, runs, and finishes in the transcript history window — with the live assistant (briefings, action items, and a running summary) right beside the transcript, a meeting-prep surface, and honest recording controls. The floating overlay is still there when you want it, but ending a meeting no longer loses your place.

The app is now fully keyboard-navigable: a real menu bar, ⌘F to find within a transcript (⌘G to step through matches), ⌘K to search meetings, and ⌘L to jump to the follow-up question box.

The live transcript also reads better and holds steadier — it splits into paragraphs at speaker turns, no longer flickers or "reloads" on every refresh, keeps its final words when you stop, and rescues quiet speech (muttered asides and trailing words) that used to drop silently.

Other improvements:

• A native macOS status panel with the Tahoe "Liquid Glass" look that follows your system Light/Dark appearance.
• Live assistant notes collapse to the most recent, with older action items and summary points one tap away.
• Custom AI endpoints: point the app at any OpenAI-compatible model, and enrich Local Mode meetings with your own provider while keeping them off the shared proxy.
• Snappier meeting search and lighter live meetings — less scroll jank, and bounded memory so multi-hour meetings stay stable.
• Fewer phantom speakers and cleaner speaker attribution on played-back media and long meetings.

Bug fixes:

• The live view no longer stalls after a speaker stops talking, or freezes on a stale tail when you stop the meeting.
• Ending a meeting from the status panel no longer leaves an empty, dimmed pane.
• Several rare crash and data-loss paths on long meetings now recover instead of failing.

This release makes long meetings far more resilient. Post-meeting processing now works from disk-backed audio instead of holding entire channels in memory, recording-time memory growth is bounded, and the app sheds caches under system memory pressure — so a multi-hour meeting no longer risks running out of memory, and several rare failure paths that could discard a finished recording now recover or preserve your audio instead. (#47)

Transcript punctuation also got a deep pass: stray periods planted at transcription window boundaries (the "inst.ructions" class) are down about 90%, duplicated sentence marks ("them..") are collapsed, a genuine question mark now survives when it competes with a spurious period, and rapid speaker handoffs fuse into one sentence less often.

Other improvements:

• Names the transcription model "remembers" but that were never actually said are now detected at window edges and flagged, so the speaker-attribution pass can veto them from context.
• Optional diagnostic file logging for support investigations, plus richer crash and reliability reports. Telemetry remains consent-gated.
• A passive watchdog now logs macOS system-audio tap dropouts (an OS bug we're tracking with Apple) for faster diagnosis.
• Short misattributed speech runs are anchored back to the correct speaker, with an acoustic veto to keep genuinely distinct voices separate.

Bug fixes:

• Stopping a meeting could rarely leave the overlay stuck on "Waiting for speech" — the meeting now completes or reports the failure.
• A meeting save could be aborted by a single malformed segment span; it now normalizes and saves.
• Transform mode no longer overwrites your clipboard selection on a failed run.
• Dictation history and usage records are no longer lost on transient write errors.
• A corrupted model download now re-downloads automatically instead of failing to load.

Your meeting can no longer disappear when you tap Stop. If transcription hits a snag while finishing up — an audio-engine hiccup or a stalled decode — the app now recovers your recording from the saved audio instead of discarding it as "no speech," and reports an empty meeting only when the recording was genuinely silent. In the rare case nothing can be recovered, the audio is preserved so it's never lost. (#47)

Other improvements:

• Onboarding now follows your system Light/Dark appearance, and the "customize shortcut" step responds correctly to clicks
• Cleaner transcripts at pauses — a filler word can no longer knock out a real word where two segments meet, and a sentence spoken across a dramatic pause stays attributed to a single speaker

Bug fixes:

• More reliable startup: a database-initialization failure now exits cleanly instead of hanging the app

The meeting overlay and status panel now follow your system appearance. Switch macOS to Light mode and the overlay switches with it — a clean light theme with readable cards and section tints — while Dark mode looks exactly as before.

The live meeting assistant works harder to put talking points in front of you the moment you need them — especially right after a remote speaker finishes a long turn and hands the floor back to you. Briefings now refresh when the conversation genuinely moves on rather than on raw talk time, and an adaptive cadence keeps short talking points from going stale on a fast-moving call. Questions surfaced to you no longer linger after they're answered, and live action items stop flickering in and out as the discussion develops.

Other improvements:

• Fewer phantom speakers after a meeting — a reworked pass dissolves grab-bag clusters built from brief backchannels and noise, and recovers speaker tails that were split off into their own bogus speaker
• More accurate speaker labeling overall, via a new content-reliability signal that down-weights low-information audio when resolving who spoke
• The primary speaker is named from meeting context, with the match confirmed when a recognized voice and name line up
• Dates in the live assistant are now localized to your region

Bug fixes:

• Fixed a freeze (beachball / 100% CPU) when scrolling a long transcript — most visible reviewing Local-Only meetings, which open straight to the transcript view. The overlay no longer fights its own auto-resize (#44)
• The app now renders correctly in Light mode — no more dark text on dark backgrounds across the overlay and windows (#40, #41, #45, #46)
• The last words of an utterance are no longer dropped at a pause — a reworked transcript merge keeps the full tail of each spoken segment

Speaker attribution is substantially more accurate this release. A reworked speaker-resolution pass cleans up the most common errors after a meeting: phantom speakers spun up from played-back audio or brief background noise are dissolved, short interjections wrongly split off a speaker are reattached, and a quiet speaker absorbed into a louder one is recovered. A brief voice that can't be placed is now parked as "Unknown" rather than inflating the participant list, and a speaker recovered from context is marked with a tentative "?" so you know it's a best guess.

Every AI feature now follows your provider choice. Document chunking, vocabulary hints, OCR, and search context route through whichever provider you've selected — cloud Gemini, your own key, or a local endpoint — so a fully local setup keeps everything on-device. Custom endpoints gain a model dropdown to pick from the models your server offers, and verifying a provider auto-adopts it across every feature selector.

Other improvements:

• Meetings show a green badge on speakers whose voice matches a saved profile, and a saved profile is backfilled onto past meetings when it's later recognized
• Edit a transcript and the summary regenerates on its own, so action items and notes stay in sync with your corrections
• Renaming, merging, or removing a speaker after a meeting applies instantly without re-running attribution
• Full keyboard navigation and focus in the overlay's editing surfaces
• Meeting notes render live as you type, support selectable markdown, and stay isolated per meeting
• Better speaker accuracy in multilingual meetings — common backchannels and fillers across Western European languages are now recognized when speakers code-switch
• A recording-consent reminder pill in the live assistant

Bug fixes:

• Paid plans are no longer metered for AI calls that don't go through the MimicScribe proxy (your own key or a local endpoint)
• The overlay re-fits its height when you switch between Summary and Transcript or move it to another screen
• Notes added or deleted from the Transcript tab now render correctly
• Closed a rare race that could drop the final words of a meeting when stopping

See exactly where the app connects. The new Network Log — in a new Activity section in Settings, alongside Transcription History — records every network request the app makes as it happens: host, purpose, and outcome, metadata only. Connection warm-up now follows your routing too: features pointed at a local endpoint or your own Gemini key warm that provider instead of the MimicScribe proxy, so an all-local setup makes no proxy connections at all, and the AI pane now labels Gemini with "your API key" when your own key handles the calls.

Meeting briefings start sooner: the opening briefing fires the moment you start a meeting, before models finish loading, so prepared context reaches the panel seconds earlier. Local and custom endpoints get a reasoning-effort control and timeouts sized for local models — larger on-device models can now finish briefings they previously timed out on — plus reliability fixes: instant failure when the server is offline, truncated-response detection, and smarter retry on rate limits.

Other improvements:

• Press ⌘F anywhere in the hub to jump to meeting search; the back mouse button, Esc, Delete, and ⌘[ now navigate consistently across the app
• Speaker attribution: a stricter merge gate stops distinct same-channel voices (e.g. two callers on a conference line) from being merged together
• Leaner AI request formats and prompt caching cut response latency for speaker attribution and summaries
• Action-item export now pre-checks only your own and unassigned items
• The MCP server now works with standard MCP clients — a framing bug had broken every client except the bundled one — and guides agents to the action items you reviewed in the app

Bug fixes:

• The assistant no longer shows doubled transcript text after a local-model hiccup
• Prepared meeting context is used for spelling and terminology precision but no longer leaks into summaries as if it were said in the meeting
• Onboarding window: drag it from anywhere; clicking buttons no longer moves the window
• Verifying a Gemini API key mid-meeting no longer leaves the keep-alive pinging the proxy

Run the AI layer locally if you want to. This release adds an OpenAI-compatible endpoint — Ollama, LM Studio, or any local server — alongside cloud Gemini and bring-your-own-key. Dictation, Transform, meeting summaries, post-meeting Q&A, and the live assistant each route independently to cloud, local, or off, from a single redesigned AI settings pane.

The live assistant is faster and easier to skim. Refreshes land on conversational boundaries instead of waiting for silence, and an on-device classifier catches commitments as they're spoken — new action items reach the panel within a few seconds. Briefings and answers are shorter, with bolding where it helps scanning. The Light plan's briefing allowance doubles to 200 a month to match the quicker cadence.

Speaker attribution got a large round of work: audio playing from your speakers no longer surfaces as a phantom speaker (an echo-cancellation timing fix), multi-speaker backchannel clutter is dissolved into the right speakers, and quiet speakers absorbed into a dominant voice are recovered.

Other improvements:

• Drag the overlay from any empty area in every mode — text still selects normally
• Transcript edits save instantly without reloading the view; speaker renames apply in place
• Custom endpoint setup verifies the model responds, not just that the server is reachable
• MCP: clearer errors, import status polling, request IDs for correlating results
• Security hardening: license and API keys moved to the Keychain, authenticated cross-process channels, tighter log hygiene

Bug fixes:

• Fixed a system-audio converter bug that silently dropped ~0.4% of remote audio — the root cause of meeting clock drift
• Transform mode verifies short selections against the clipboard, fixing bogus one-character captures in Google Docs
• Assistant panel no longer jiggles, clips its last line, or flashes a loading dim on reopen
• Played media no longer generates spurious action items, and the action-item list no longer drops or duplicates entries between refreshes

AI meeting summaries are now free on every tier — with separated, named speakers and transcript correction, and a generous daily fair-use limit. Paid plans now center on the live in-meeting assistant and bring-your-own-key. File transcription has its own daily limit, separate from meeting summaries.

Summary views replace meeting types, and you can now build your own. The four built-in views — Simple, Walkthrough, By Participant, and Decisions & Open Questions — sit alongside any custom view you create. Set a default in Preferences, or switch the view from the post-meeting summary header. Switching a view restyles only the narrative; the overview, action items, and speaker transcript stay put.

Other improvements:

• Reworked the post-meeting summary header: an inline "View:" switcher in the metadata row, a cleaner title and speaker layout, and steadier spacing.
• The pre-meeting context form now opens straight to a notes field instead of a row of template chips.
• Evened out the spacing around the Q&A divider.

Bug fixes:

• Switching summary views no longer dims or reloads the speaker transcript.
• The overview and action items stay stable across view switches instead of flashing a reload.
• Fixed a one-word phantom speaker (for example, "go.") being split off from the person who actually said it.

This is a focused follow-up to the speaker-identity work in rc.5, mostly making voice recognition stick more reliably. You can now be enrolled from a single long turn — a continuous solo talker is merged into one segment, which the old two-segment enrollment bar rejected no matter how long they spoke. Saving a global voice profile now always creates a durable, verified profile instead of occasionally leaving a "(You)" label with no recognized badge, and that badge persists after you leave and reopen a meeting.

Other improvements:

• When two speakers are merged, both voices feed the saved profile instead of dropping the shorter cluster, so recognition holds up better after a correction
• Transform mode is now marked Beta in Settings, with a heads-up that it may move to an overlay that previews edits before applying

Speaker separation is substantially more accurate. A rebuilt sentence-level pipeline, smarter cluster merging, and new guards against bleed and end-of-meeting artifacts mean far fewer split, duplicated, or phantom speakers, and the attribution prompt was rewritten for cleaner who-said-what. The app also recognizes you by voice across meetings: confirm "Is this you?" once and the (You) badge follows you forward and backward through your history, with inline renaming and one saved voice profile per person.

Much of that accuracy comes from a new echo canceller. Early builds used DTLN, which never fully removed the echo and left distracting artifacts; Apple's built-in voice processing cancelled cleanly but ducked the meeting audio itself, quieting the conversation you were trying to capture. RC5 switches to LocalVQE, an on-device neural model that runs on copies of the audio — it cancels more completely than DTLN without the ducking, so what you hear stays untouched while the mic is cleaned before transcription. It downloads in-app on first use and, with a stack of new mic-bleed defenses, clears most of the phantom "duplicate" speakers caused by a remote voice leaking into your mic.

The live meeting assistant gained an in-meeting Q&A bar and a running catch-up feed of summary points that updates and retracts itself as the conversation moves. Post-meeting summaries now adapt their structure to the meeting, lead with a tight overview and tiered action items, and can be reshaped with view lenses like Decisions & Open Questions. Every edit you make is reversible with multi-level undo and a one-click restore-to-original.

Pricing is now three tiers: a Free plan with daily on-device voice typing plus a starter allowance of AI features, a Light plan for regular meeting use, and Unlimited. Bring-your-own-key bypasses caps on any paid tier, and the in-app subscription panel was redesigned around the new tiers.

Other improvements:

• Local Mode is enforced end-to-end — across follow-up Q&A, search, and the MCP tools — with on-device speaker recognition so a private meeting stays fully on your machine
• The menu bar shows "Capturing audio" with elapsed time while a meeting records
• Faster, smoother transcripts: the meeting transcript is virtualized and rendered natively, with per-turn selection and a lighter markdown path
• Keyboard navigation across the meeting hub and summary; standard macOS pointer cursors throughout the overlay
• Accessibility: opaque, bordered controls and high-contrast treatment under Reduce Transparency / Increase Contrast
• Per-turn speaker reassignment directly in the summary, with clearer "Reassign to " labels
• The entire AI pipeline moved to Gemini 3.1 Flash-Lite for faster, lower-cost responses
• File imports run as explicit MCP tools with determinate progress and ETA
• Summaries keep a dedicated section for notable personal updates and split long monologues into readable paragraphs
• Security: Sparkle updater bumped to 2.9.2; server logs scrub email PII

Bug fixes:

• Fixed two cases where a meeting could hang at "Identifying speakers," plus a database-init deadlock on startup
• Hardened a range of crash and concurrency foot-guns across the recording, audio, and fusion paths
• Interrupted recordings are recovered and surfaced with a one-time launch alert instead of being lost
• Free-tier usage meter stays in sync with the server; billing charges once per action rather than per network call
• Cold-start meetings no longer cancel on hotkey release or show a false "Live transcript paused" banner
• Post-meeting Q&A no longer echoes the transcript when a question is unclear or unrelated

This release candidate is a substantial step up in voice-based identity, diarization accuracy, and meeting-assistant quality. The (You) badge is now gated on voice match and survives across meetings via stable profile-UUID colors, with tap-to-rename badges, an inline "Is this you?" pill, and name edits that cascade across history. Speaker recognition uses multi-centroid profiles for cold-start matching (−42% relative EER at 60 seconds of voice), and a new acoustic boundary-snap pass tightened diarization edges by another −0.76pp DER on top of a re-shipped attribution prompt.

Briefings and summaries got tighter prompts and stronger guardrails: TALKING POINTS is slimmer, USER_CONTEXT no longer bleeds product nouns into off-topic meetings, action-item due dates resolve against meeting timezone with a dual-clock display, and past-due items are preserved across refreshes. Dictation moved to a more interpretive prompt validated against a 37-case benchmark, and the vocabulary system now phonetically gates spelling hints (Double Metaphone), supports n-gram compound terms, and classifies user-supplied terms by domain so non-tech vocabularies aren't biased toward developer jargon.

File imports gained drag-drop, Open With support, and in-flight progress UX. Enrichment failures now surface a reason and a raw-transcript banner instead of failing silently — including the Gemini SAFETY block path, which was previously masked by content filters. The Pro tier is renamed to Unlimited everywhere.

Other improvements:

• History view collapses to two filters (Meetings / Dictation & Instructions); "Use as Context" removed from the multi-select bar; copy-transcript button at bottom of detail
• Status panel shows an empty-state hotkey card and auto-shrinks
• Single-level undo for voice-driven meeting corrections (Cmd+Z)
• Meeting Q&A: per-row copy / revert / delete actions; past corrections shown inline in the history-panel summary; classifier now extracts speaker-rename pairs cleanly
• Overlay: skeleton loader for the first briefing, peek briefing without stealing focus, first-run empty-state card replaces the coaching popover, transform-independent toggle, redesigned waveform, press feedback on meeting-hub cards
• Transcript and summary panels rendered with vendored Textual for cross-paragraph selection
• Network calls now use jittered exponential backoff
• Voice profile saved before renaming segments so the green check lights up reliably
• Speakers in file imports now get voice embeddings; profile centroids un-freeze and recompute from samples; merging two profiles preserves embeddings
• Outlier voice samples are trimmed before centroid averaging
• Three-level speaker identity in the database with proper merge cascading and orphan sweep
• Website: GDPR & CCPA trust link under the hero CTA, inline monthly/yearly billing toggle on the pricing card, copy refresh ("speaker separation" instead of "diarization", "Library" instead of "Knowledge Base")

Bug fixes:

• Cmd+D delivery is reliable across prewarm, escape cancellation, and stop-during-startup races; "Listening…" holds through finalize and cancels cleanly on view teardown
• Gesture state resets when recording ends externally
• Meeting hub no longer claims "Processing" when a meeting is only queued
• Briefing pinned at first streaming chunk to stop the floor from jumping
• Overlay size jumps eliminated on start-meeting bootstrap and briefing refresh
• Hub reopen no longer flashes; template chip first-click is reliable
• Meeting corrections preserve verified speakers and dim the summary during regen
• Empty LLM-output segments are dropped during fusion; orphan meeting_speakers are swept
• Fallback speaker labels shortened to "Mic N" / "Remote N"; falls back to "Speaker N" when display_name resolves empty
• Email correspondence in Insert mode now gets paragraph breaks
• Floating panel toolbar view tree stays stable so the hub→summary morph fires
• Meeting-hold pill aligned with the waveform capsule
• "Reveal in History" wired in the unified overlay summary
• Brighter assistant loading labels for legibility
• Audio level overlay hidden on every recording cleanup path
• Persistent billing gate banner with rate-limit attribution

The meeting overlay got a significant polish pass. The briefing panel now streams in as it generates and auto-refreshes when you pause after new speech, so the view stays current without you reaching for the refresh hotkey. The whole panel is draggable, speaker colors are unique per speaker (blue reserved for You), and chrome across every window now shares a unified look. Transform mode gained writing-sample support so it can better match your voice, plus anti-injection guardrails that prevent your source text from being treated as instructions.

Other improvements:

• Briefing preserves durable facts in the summary across refreshes instead of rewriting it each time
• Briefing template quotes captured questions verbatim
• New setting to disable briefing auto-refresh if you prefer manual control
• Settings > Dictation gained a short intro; meeting shortcut descriptions clarify global vs window scope
• Billing migrated to Stripe; referral program added for existing customers

Bug fixes:

• Audio-level capsule follows your active app window after the overlay resigns key
• Floating panel no longer swallows global keystrokes or gets stomped by alt-tab
• Click-anywhere-to-key restored on the floating panel
• Overlay follows to whichever screen you interact with it on
• Meeting assistant speech tracking uses sustained VAD state (fewer false triggers)
• Concurrent live-transcript requests are deduped instead of racing

This release candidate focuses on reliability and responsiveness in long meetings. Gemini responses now stream end-to-end for attribution and summarization, eliminating Cloudflare timeouts that previously hit 90+ minute recordings. Diarization accuracy improves with word-cut boundary repair and re-tuned segmentation — measured +2.1pp speaker attribution on AMI.

The overlay gained a substantial stability pass: in-flight meetings can be reopened from the Hub while still processing, the assistant↔hub transition recovers cleanly from start errors, the summary view transitions correctly on stop, and the end-meeting flow is unified across live and deferred paths. Processing status throughout the UI is now truthful — the perpetual "Processing" badge has been replaced with actionable failure states, and spinners no longer appear when nothing is active.

Per-meeting summary templates (Discovery, Sales, Interview, Standup, Customer, Presenting) now ship with tuned instructions honed against an 18-scenario benchmark. MCP agents can stage a template before a meeting via set_meeting_context with template_id, or discover options with the new list_templates tool.

Briefing and summary prompts have been re-tuned from A/B data to improve requirements gathering, resurfacing of prepared context, and section routing. Transcript history search now supports voice input for follow-up queries, with skeleton loaders and unified result styling.

Other improvements:

• Streaming Gemini calls for long attribution and summary prompts
• Word-cut boundary repair in diarization; step=0.16, attribution temperature tuned to 0.1
• Per-template summary instructions for all six built-in meeting types
• MCP list_templates tool and template_id parameter on set_meeting_context
• Voice input for transcript history search follow-ups
• Copied meeting summaries now include user notes
• macOS 15.0 set as the official minimum deployment target
• New cost-comparison benchmark page on the website

Bug fixes:

• Perpetual "Processing" badge replaced with actionable failure states
• Spinners no longer shown when enrichment is idle
• Meetings still enriching can be reopened from the Hub
• Assistant and Hub transitions recover from start-error states
• Summary view transitions correctly after stopping a meeting
• Briefing state wiped on meeting stop, not only on start
• Proxy allowlist now matches camelCase :streamGenerateContent
• Explicit transcript content preserved during refinement safety passes
• No-speech meetings handled consistently across live and deferred paths
• Billing gating no longer triggers during upgrades or transient errors
• Assistant header drag works on first click (SwiftUI gesture)
• Speech-pace alert sensitivity reduced to cut false positives
• Enrichment badges use larger caption font for readability

This release candidate sharpens the meeting assistant around discovery and requirements gathering, with substantial prompt and context work to reduce hallucinations, improve resurfacing of prepared context, and keep pace in long meetings through preemptive transcript compaction. A new Presenting template lets the assistant track coverage during demos and walkthroughs instead of nudging goals.

Context retrieval has been rebuilt on MiniLM embeddings with conversational search-phrase chunking, multi-query retrieval, and paragraph-split indexing — substantially improving recall on prepared reference documents and past meetings. Reference documents now accept a wider range of formats via drag-and-drop, including images through Gemini OCR.

The meeting Q&A experience now supports three distinct intents: questions answered against the transcript, user notes captured inline, and corrections that regenerate the summary in place. Conversations persist across reopens, user notes survive summary regeneration, and the classifier routes each turn without the main LLM having to guess.

Meeting recording is more resilient: state is preserved across quit, sleep, and crashes; billing credits are no longer rolled back on transient refinement errors; and code-switching is supported in primarily-English meetings. Silero VAD is now used for device-independent speech detection, powering the refresh badge and idle monitor.

Other improvements:

• New dedicated hotkey to refresh the meeting assistant
• Insert-mode history shows the original transcript alongside the refined text
• Dual-channel diarization runs concurrently for faster meeting finalization
• Meeting summaries gain a unified glass overflow menu and smoother processing-to-summary transitions
• Onboarding and settings panels adopt macOS Tahoe glass styling throughout
• Gemini proxy hardened with model allowlist, device ID validation, and a licensed-user monthly cap
• Prompt injection mitigations added across all LLM prompt sites
• Action items and summary prompts no longer extract from user/reference context
• Speaker badges show consistent colors across overlay, history, and search

Bug fixes:

• Bare-key global hotkeys no longer swallow system-wide input
• Meeting assistant hotkey no longer gets permanently stuck
• Per-meeting reference docs now display correctly instead of the current global state
• Stale LLM stats cleared on metadata and correction turns
• Briefing no longer resolves relative dates (keeps ISO in action items)
• RAG pipeline triggers for MCP-set context and during-recording edits
• Escape and Cmd+W dismiss all overlay modes
• Accessibility permission state refreshes in the dictation UI
• Overlay header drag works on first mouse click
• First heading no longer shows a redundant divider

This release introduces the Meeting Hub — a redesigned overlay with one-tap meeting controls, live VU meters, and integrated meeting prep. The entire UI has been refreshed with macOS Tahoe-inspired glass styling, bringing a modern look to overlays, toolbars, and interactive elements.

Meeting assistant now auto-refreshes intelligently based on conversation activity, and briefings persist across dismiss/reopen cycles. Idle recording detection prompts you when no speech is detected, helping avoid forgotten recordings. Instruction mode has been streamlined to direct-paste with session history.

For privacy-conscious users, Local Mode (previously Offline Mode) is more prominent, and a new Bring Your Own Key option lets licensed users connect their own Gemini API key. The privacy page has been overhauled with clearer language around permissions and data handling, and the Gemini proxy is now open source.

Other improvements:

• Unified voice typing quota for dictation and transform modes
• Raised free-tier limits for beta
• Reference documents now route through the embedding pipeline for better retrieval
• Single Gemini call for context processing with concurrent chunk handling
• Speaker attribution uses acoustic hints and conservative bias for better accuracy
• Echo leakage detection in diarization

Bug fixes:

• Fixed embedding model load race when multiple requests arrive simultaneously
• Empty meetings no longer saved to database
• Q&A responses now route to the correct view model
• Fixed diarization crash on duplicate speaker mapping keys
• Fixed instruction mode overlay race conditions during deferred upgrade
• Preserved contractions in dictation refinement
• Reduced over-eager vocabulary injection in refinement prompts

The meeting assistant now surfaces interpersonal awareness more accurately. Acoustic signal annotations — overlapping speech, laughter, crosstalk, and similar events — are detected in the transcript and sent to Gemini alongside the spoken content. This gives the assistant richer context for understanding conversation dynamics, not just what was said.

Onboarding has been redesigned with a hold-to-practice interaction and a coaching overlay that guides new users through their first recording. The meeting assistant overlay received a round of UI polish: no more width jumps on load, smoother transcript transitions, and clear empty states when no speech is detected.

Other improvements:

• AI agents can now manage reference documents via the new set_context_source MCP tool — add, update, or remove context that informs your meetings
• Meeting assistant transcript auto-refreshes in offline and billing-blocked modes
• File imports now go through the full speaker attribution pipeline
• Meeting summaries lead with decisions and use a bullet-only format for faster scanning

Bug fixes:

• Hotkey monitors re-register correctly after a binding change
• Microphone permission prompt no longer appears prematurely on first launch
• Dictation enrichment timeout increased to 3s to reduce premature cutoffs
• Meetings with no detected speech show a clear empty state
• Fixed checkout API token mismatch and license key selection

Add reference documents to your meetings — upload agendas, briefs, or prior notes and they're chunked, embedded on-device, and retrieved during the meeting to give your assistant grounded context. Indicators show when documents are loaded and a test mode verifies retrieval quality.

MimicScribe now includes an MCP server that connects your AI coding agent to your meeting history. Eight tools — search meetings, get transcripts and summaries, query action items, set meeting context, manage reference documents, and import audio files — work with Claude Code, Cursor, Zed, and other MCP-compatible clients. The server sends real-time notifications when meetings are enriched, imported, deleted, or merged. (#27)

Batch file import lets you queue audio and video files for transcription with speaker diarization — drag files in, add them from the context menu, or use the MCP import_audio_file tool. A new queue display tracks processing status.

Meetings can now pull attendee names from your calendar automatically, and a new per-meeting Offline Mode disables all cloud AI while keeping on-device transcription running — you can retroactively generate a summary later. (#31)

Other improvements:

• Unified meeting overlays into a single panel with Control+Space toggle
• Meeting assistant auto-refreshes its briefing when a pause in conversation is detected (#23)
• Custom vocabulary for specialized words and names that ASR and LLM should recognize (#22)
• Inverse text normalization — numbers, dates, phone numbers, and times are auto-formatted in transcriptions (#24)
• Configurable dictation timeout with raw text fallback (default 2s)
• Double-tap recall now works in dictation mode (reopens last result)
• Cmd+V paste from instruction overlay, mouse-cursor screen selection
• Instruction mode matches the input format instead of defaulting to Markdown
• Cmd+W closes all overlay windows
• Error sounds and visual feedback for dictation/instruction failures
• Accessibility permission is now optional (only required for voice typing)
• Daily billing caps for assistant and follow-up queries (10/day), file transcription cap, briefings counter in free-tier panel
• Two-column history panel with speaker colors

Bug fixes:

• Fixed hotkeys not working after onboarding completion (#15)
• Fixed overlays jumping during resize (now anchored from top edge)
• Fixed system shortcut conflicts not being detected during onboarding
• Fixed instruction mode stripping metadata wrapper from selection context
• Fixed daily billing limits resetting at UTC instead of local midnight
• Fixed dismiss-first tap behavior and minimize-aware overlay visibility
• Suppressed network calls in offline mode and billing-blocked recordings
• Fixed clipboard fallback for Zed and GPU-rendered editors

Dictation now adapts to the app you're using. Per-app profiles let you set different transcription styles — casual for messaging apps, polished for documents — and automatically switch based on the active window, with browser URL matching for web apps like Google Docs or Slack. (#12)

Meeting overlays are now hidden from screen sharing by default, so your AI assistant stays private during Zoom, Meet, and Teams calls. (#25) A new set of quick-fill context templates lets you describe your meeting in one click, and you can now add or edit meeting context after recording has started. (#17) Meeting summaries support custom format instructions — configure once in settings to get the output style you prefer. (#13)

Other improvements:

• Inline speaker renaming in post-meeting transcripts with automatic profile saving (#19)
• Audio level overlay follows your active window across screens (#14)
• Meeting assistant overlay remembers its position between uses
• Replaced trial token system with a permanent free tier
• Added Gemini 2.5 Flash Lite model and thinking level selector
• Hold-to-talk no longer triggers from held keys (key repeat filtering)
• Human-readable dates in meeting assistant briefings

Bug fixes:

• Fixed meeting summary window appearing over full-screen apps (now bounces Dock icon instead of stealing focus) (#18)
• Fixed overlays showing on the wrong screen when the frontmost app is full-screen
• Fixed speaker label colors resetting after a rename

Meeting summaries now open in a proper macOS window with a title bar, traffic lights, and keyboard shortcuts (Cmd+W to close, Shift+Cmd+C to copy). The window auto-sizes to fit content — starting compact during processing stages, then growing when the summary arrives (KVO-based content fitting, capped at 80% screen height). You can ask follow-up questions directly in the summary window, with full conversation history preserved between questions.

The menu bar icon has been updated to the new bird logo, and the overlay system has been migrated to @Observable for per-property tracking, reducing unnecessary SwiftUI re-renders.

Other improvements:

• About You field now accepts longer text and includes writing style guidance
• Selection capture works in canvas-based editors like Google Docs (AX + clipboard fallback)
• Rich text paste no longer inserts spurious blank lines (block-aware markdown parser)

Bug fixes:

• Fixed a crash when reusing the meeting summary window (Auto Layout constraint loop)
• Fixed meeting recording getting stuck in "stopping" state when no active worker
• Fixed the last few seconds of audio being lost when stopping a meeting (DTLN buffer not flushed before finalization)
• Fixed meeting summary window appearing off-screen on smaller displays
• Fixed meeting assistant panel appearing on the wrong screen in multi-monitor setups
• Fixed login item re-registering on every app launch (SMAppService persists across launches)
• Fixed onboarding window not behaving like a standard macOS window (activation policy)
• Fixed stale selection badges and synchronous database reads in the status panel
• Fixed phantom Q&A entries appearing in meeting history
• Fixed window drag gesture using wrong window reference

Stability release focused on fixing the crashes that made beta.2 unusable on first launch. The DTLN echo cancellation model now loads correctly from the app bundle (SPM Bundle.module bypass), and the loading overlay no longer blocks your first dictation after onboarding.

Onboarding also got visual polish: an accessibility settings mockup shows exactly which permissions to grant, and interactive keyboard visualizations on the hotkey pages let you preview each shortcut before committing to it.

Onboarding has been completely redesigned into a 4-page flow with per-shortcut setup, so each recording mode is introduced and configured on its own page. Privacy policy, terms of service, and opt-in analytics are now presented during setup rather than buried in settings.

A new Privacy settings panel lets you control what analytics and crash logs are sent to the developer. Dictation mode now rewrites more aggressively—cleaning up contradictions, long pauses, and restarts into clear, concise language. Meetings now queue for offline enrichment when there's no network connection, processing automatically when connectivity returns (persistent SQLite queue).

Other improvements:

• Launch at login enabled by default for new users (SMAppService)
• Meeting assistant default shortcut changed to Control+Space
• Monospaced font for keyboard shortcuts display
• Consistent spacing hierarchy in settings window
• Assistant usage hint added to meeting start panel

Bug fixes:

• Fixed event tap re-enable loop when accessibility permission is revoked
• Fixed database migration for missing enrichment columns (reused v10 migration)
• Fixed volume limiting default to 10% for new installs
• Fixed audio feedback sounds cutting off and engine lifecycle issues

First public beta of MimicScribe — a macOS menu bar app for speech-to-text. Transcription is powered by NVIDIA's Parakeet TDT 0.6B model compiled to CoreML, so your audio never leaves your Mac. Text refinement, meeting summaries, and follow-up questions use Gemini.

Meeting mode (Cmd+Shift+Alt+Space) captures system audio and microphone simultaneously, producing a diarized transcript with speaker labels and an AI-powered summary. Speaker diarization uses pyannote community-1 models with a Gemini pass to fix attribution errors. DTLN neural echo cancellation separates your voice from system audio. A real-time meeting assistant lets you ask questions about the conversation while it's still happening.

Insert mode (Alt+Space) transcribes speech and pastes it at your cursor. Instruction mode (Shift+Alt+Space) sends voice commands to Gemini with your selected text as context, displaying results in a dark glass overlay.

Other improvements:

• Dark glass onboarding flow with card-based layout
• Sparkle auto-updates with EdDSA signing
• Custom CGEventTap hotkey system (replacing KeyboardShortcuts library)
• Redesigned status panel trial card with purchase and activate links

Bug fixes:

• Fixed 3-second hang on app quit (main-thread deadlock in audio teardown)
• Fixed DTLN model crash in .app bundles (SPM bundle resolution)
• Fixed onboarding accessibility setup not guiding users through permissions