Roadmap
Dictation is
the entry point.
Everything below is either running on machines today or actively being built. The version beside each entry says which build it lives in: stable is what a download gives you, beta and dev run ahead of it. Where we have measured numbers, they’re here — including the ones that still need work.
Dictation that disappears
The core loop, on every machine running the stable build: hold a key, talk, get clean text in any app — with transcription and cleanup both running on your own Mac.
- Streaming transcription as you speak (Parakeet, local)
- Smart paste routing for native, web, and terminal targets
- Deterministic cleanup, ITN, and a custom glossary
- Insights dashboard with WPM, streaks, and usage broken out by app
- Nearby Macs, so one machine can transcribe for another — opt-in, TLS-paired, and off until you turn it on
- Signed, notarized, and updating itself via Sparkle
The assistant, and what it can reach
Hold fn + control instead of the dictation key and the same voice files a note, sets a reminder, or asks a question of the tools you've connected. Live on the beta channel now, and deliberately dark on stable until it has been proven there.
- Eleven connectors reading for real: Apple, Google, and Exchange calendars, Gmail, Drive, Slack, Notion, Linear, GitHub, Zoom, Asana
- Writes — a Slack message, a calendar event — only ever behind an approval card you see first
- Notes and reminders by voice, each note keeping the recording it came from
- Answers spoken back, on the same on-device model as smart cleanup
- Regulated Mode, enforced at each network egress rather than promised in a document
- A release that explains itself: what's new, shown once after an update
A coding agent in the notch
The notch already takes your voice and shows you an answer, so pointing it at an agent already running on your Mac costs almost nothing. Start a turn by talking, watch the tool calls land, and answer a permission ask without leaving the app you're in.
- One customizable key sends a spoken prompt to a live agent session
- Permission asks answered from the band, with Ask, Auto, or Plan per session
- A finished turn opens as a card that renders the markdown real replies are made of
- Client only — nothing is installed or supervised, and with no agent running the surface stays absent
Better ears, softer edges
The evaluation engine says clean speech is solved at 3.4% word error rate, and that noise and the audio Bluetooth serves up in call mode are the real bottlenecks. Those numbers have not moved yet, so they stay here.
- Accuracy in a noisy room (23.5% WER today, the biggest open gap)
- Bluetooth headsets (16.3% WER over HFP, and everyone dictates on AirPods)
- Smart cleanup out of the settings and into the default path
Everything you said, in one private place
Dictation is the entry point, not the destination. The same local pipeline that hears a sentence can capture a meeting — and then let you ask questions about the whole of it.
- iPhone as a wireless mic and dictation keyboard
- Local meeting capture for calls on your Mac
- One searchable memory across dictations, notes, and meetings
Want a say in the order?
Early users decide what gets pulled forward. Grab a build, use it for a week, and tell us what got in your way.