Our own product, on our own P&L. A native desktop app for Mac and Windows that dictates into any text field, records both sides of a call, and turns it into summaries, action items and agent routines, with transcription running locally.
Meeting bots join the call as a stranger, and cloud notetakers ship every word to someone else's server. For lawyers, founders and anyone under NDA, that rules the category out entirely.
A notetaker with no bot and no upload by default. It captures the call from the operating system itself, transcribes on the user's CPU, and keeps the result as plain files the user owns and can open in anything.
A workspace: push-to-talk dictation, meeting and screen recording, a searchable vault, voice commands, and routines that run on a schedule against Gmail, Outlook and Notion, exposed to Claude and Cursor over MCP.
Shipped release by release, up to v1.11. We own the roadmap, the support queue and the release train, so every one of these runs on our time and our money.
Hold a shortcut, speak, release. The text is typed into whatever field has focus, with word-by-word streaming and a cleanup pass.
Your mic and the call's audio are recorded separately, so the transcript knows which words were yours. Screen video is optional.
Action items come with an owner and a priority, written by Claude through our cloud, the user's own key, or a model running locally.
Everything is saved as plain Markdown files on the user's disk. ⌘K search across it all, a one-click Obsidian plugin, and nothing locked in.
Scheduled agents that read the meetings and act: pre-meeting briefs, follow-up drafts, and commitments closed once they're done.
Claude and Cursor can read the user's meetings, people and action items directly, without the app having to be open.
The meeting pipeline enforces one rule: the mic is the user and nobody else, and system audio is everyone else. Each stage exists because an earlier version broke it.
Your mic and the call’s audio are recorded as two separate tracks, straight from the operating system. No bot joins.
The two tracks are lined up precisely. A tenth of a second out, and echo removal stops working.
Echo, background hum and uneven volume are removed, so both sides sound clear.
Silence is skipped, then one of 17 on-device models, or the cloud if the user prefers, writes the transcript.
Summaries, chapters and action items land in the user’s own files, ready for Claude or Cursor to query.
The problems that decide whether a product like this ships at all, and what solving each one means for the people using it.
Laptop speakers leak the other side of the call into the mic, so the transcript hears everyone twice.
We record your mic and the call's audio as separate tracks, line them up, and strip the echo out, without cutting into your own speech when two people talk at once. Two earlier approaches did one or the other; we shipped only when it did both.
Clean attribution: your words are yours, everyone else's are theirs
The usual way to record a Mac's audio asks for Screen Recording access, and many people say no.
Audio-only meetings use a narrower macOS audio permission, and screen access is only requested when someone actually wants video. On Windows the same capture needs no permission at all.
No Screen Recording prompt for audio-only meetings
macOS can report a permission as granted when it isn't, and the app then quietly fails.
Hyprcore checks permissions the way the OS itself stores them rather than trusting a cached answer, and watches for the grant instead of waiting on a dialog that can hang.
First run ends with your own words coming back, not a spinner
A recorder that steals focus mid-call, or shows up in the deck you're presenting, gets uninstalled.
The recorder, banner and dictation pill are native floating panels that respond to a click without pulling focus from the meeting, and are hidden from screen capture.
Present, take notes and record in the same call
Some users want nothing to leave the laptop, others want the best cloud model, and the business needs a paid path.
Every transcription and AI call goes through one router the user controls: local only, cloud only, or somewhere between. The default is our own metered cloud, tied to their plan; their own API keys and local models are always an option.
Privacy for users, a revenue path for the business
Accuracy is the first thing a bad review complains about, and you can't protect what you don't measure.
Transcription and speaker-labelling accuracy are scored automatically against reference recordings, so a model or code change that makes things worse is caught before release, not by users.
Word and speaker error rates tracked on every change
Chosen per layer: native where the operating system matters, managed where it doesn't, and always with a local option for the AI.
DextegoClient build · AI sales coachingA full AI sales-coaching platform: live in-call guidance, roleplay with voice coaches, buyer profiling and post-call scorecards, across a web app, API, Chrome extension and desktop app.
Point at the case study closest to your problem, or just describe it. You'll talk to an engineer from the bench that built these, not a salesperson.