Summary
One Natively key runs everything — AI, speech and search move to Natively API, a diagnostic report option, a trial that starts on more machines, cleaner local speech, and Pro on Windows.
What's New
- One Natively key — a pack or plan key moves AI, speech-to-text, the reranker, embeddings and search to Natively API. Clearing the key, or the server refusing it, puts your settings back.
- Send diagnostic report — when Natively can't reach its servers, it says why and offers to send a report with a masked log tail. Opt-in, in Settings → Natively API.
- Connection help — when Natively can't be reached, the app works out the likely cause and tells you what to try, instead of a generic failure.
Improvements
- Held pack keys — a pack key you already hold is used first, and your plan carries on afterwards.
- Natively key on update — existing customers who had their own AI or speech key have it switched to Natively API when they update, once per key. Their own key stays saved, and it comes back when the Natively key is cleared. A plan key is checked with the server first, an allowance already used up stays where it is, and the switch waits for a running meeting to end.
- Re-index on switch — if you pinned an embedding provider, your notes are re-indexed once on Natively, and again if your settings are put back. The reranker and embeddings stay on your device when Settings keep reference files local.
- Web search — with a Natively key, search uses Natively API first, and your own Tavily key answers once the plan's research runs out.
- Pro box — a pack key pasted into the Pro licence box is saved as your Natively key, so the AI uses it too. A second, different pack key while one is in use is refused with a message, and your current licence stays.
- Trial start — when the native module can't give a device ID, it is read from the system (macOS, Windows registry or Linux machine-id) and gives the same ID, so the free trial starts on machines where it was refused.
- Pack and plan changes — a pack's Pro goes with its key, a plan that ends mid-pack keeps its Pro switches, and yearly plans move up in place.
- Yearly prices — the yearly price on plan cards takes turns with the monthly one, and every plan card is as tall as the packs.
- Crash details — the crash report says where the run crashed, how far startup got and what it last did.
- Auto Answer sharing — the consent text now says that typed chat and AI answers are shared along with transcripts when you share them.
Fixes
- Windows Pro — Pro now activates on Windows: the app trusts only usable roots from the system certificate store (#759).
- Local speech — Nemotron's Auto Detect no longer returns empty or gibberish text in Spanish, French, German or Hindi.
- Local speech — invented phrases such as "Thanks for watching!", sound labels such as "(upbeat music)", and interjection-only lines are dropped. Short replies like "yeah" and "okay" are kept; "uh-huh" and "mm-hmm" are dropped too.
- Local speech — Whisper's silence checks now run, so non-speech audio is dropped instead of transcribed. On slower models, final text no longer lags 12–25 s behind the speaker.
- Windows taskbar — the taskbar shows the Natively or disguise icon, not Electron's default icon.
- Network errors — a failed Pro or licence request now says why it failed, instead of only "fetch failed".
- Search embedder — the on-device embedder picks up a new Natively key as soon as you change it.
- Launcher — the search box no longer opens by itself when ⌘B brings the window back.
- Funnel counts — answer-button presses are counted by what came back, so answers that returned nothing show up in the funnel report.
Technical
- Natively switch: electron/services/packs/packNativelySwitch.ts, wired in ipcHandlers.ts. A plan key is checked with /v1/usage first; a pack that has ended is put back once it looks gone on two runs.
- Local STT: Whisper's no-speech, repetition and log-probability guards are computed in the worker (whisperSilenceGuard.ts), since transformers.js 3.8.1 ignores the library thresholds.
- Connectivity diagnosis, probes and a watcher run in the main process; the certificate-store undo is taken back when it does not fix the connection.
- natively-api submodule at 1b0f100 (the live API: Voyage hedge and review fixes, plus the press-count and device-ID fields); premium submodule at 369b379 (PR #14, licence network cause). Deploy the API before shipping this build.
- Trial device ID: electron/services/machineIdFallback.ts, same SHA-256 as the native module, no host-name fallback. The Windows registry read runs
reg.exeonce, hidden, only after the native read fails. - Mac DMG builds retry the create-dmg Finder step and notarize the arm64 DMG after the app is built.
- Electron 43.7.7 and Node 22.13 or later, unchanged.
macOS Installation (Signed Build)
Download the correct architecture .dmg or .zip file for your device (Apple Silicon or Intel).
- For Apple Silicon (M1/M2/M3/M4): Download the
arm64build. - For Intel Macs: Download the
x64build.
Open the downloaded file, drag Natively to your Applications folder, and launch it.
⚠️Windows Installation (Unsigned Build)
When running the installer on Windows, you might see a "Windows protected your PC" warning from Microsoft Defender SmartScreen saying it prevented an unrecognized app from starting.
Since this is an unsigned build, this is expected. You can safely ignore it by clicking More info and then Run anyway.
See CHANGELOG.md for the detailed list of changes.