EST. 2026 · BERLIN

We didn't make recordings safer.
We stopped making them.

Every other notetaker asks you to trust that it deletes your recordings. Fazit removes the thing you would have to trust: your call is transcribed on your Mac, as it happens, and the audio is never written to disk in the first place. There is no recording to leak, subpoena or regret.

Your account, licence and updates use our servers, and product analytics are opt-in. That lane is small, separate from your calls, and written down on Legal.

What Fazit does now

It runs the whole call, end to end.

None of it asks anything of you mid-conversation. That is the point: the work happens where you are not looking.

Starts and ends itself

Fazit notices a call beginning in any app and captures it. When you leave the meeting it writes the note. Muting yourself never ends a session, because it watches the call rather than listening for quiet.

The note, not a transcript

Summary, decisions, action items and a follow-up email in your client's own tone, written to Obsidian or Apple Notes as a plain file you own. The transcript rides along underneath, folded away.

Live meeting cost

What the call is costing while you are still in it, from rates you set or real ones read from your calendar and payroll. The figure lands in the note, where it is still true next week.

Hold fn to dictate

Anywhere on the Mac: hold the key, speak, let go, and your words arrive in whatever field had focus, punctuated and de-stuttered, never rewritten.

Every call, still here

A History tab for everything Fazit has written, with search. It reads the notes already in your vault rather than keeping a second copy, so deleting one there removes it here too.

On this Mac

Speech recognition and the model that writes your note both run locally. For the note itself, the only server involved is localhost.

Why we exist

Client work in law, finance, consulting and therapy runs on conversations that have to stay private. Cloud notetakers put those conversations on someone else's servers: the audio is uploaded, a recording exists, and a third party has joined a confidence that was never theirs to hold.

Fazit removes the recording instead of protecting it. Transcription and note-writing run on your Apple Silicon Mac, the audio lives in memory for seconds, and what remains afterwards is a text file you own.

Everything else Fazit does is built on top of that guarantee.

Architecture you can audit

Five facts, each enforced in code rather than promised in a policy. The full write-up, with the code paths behind each one, is on the security page.

  • Two physical audio streams, so You and Them attribution is exact rather than guessed.
  • A RAM-only ring buffer for audio. A write-to-disk path does not exist in the code.
  • Parakeet through CoreML for speech, a local model through Ollama for the note.
  • Notes written straight into the Obsidian vault or folder you choose.
  • Every note declares audio_retained: false, and source access for security review is available on request.
Who uses it

Built for work that cannot afford a leak.

Coaching and consulting

Session notes and follow-ups where full presence is the job, and clients never agreed to a recording.

Government and public sector

Policy, inter-agency and citizen-facing work where handing conversations to a transcript vendor is a non-starter.

Legal, finance and regulated

Attorney-client, M&A, board and compliance conversations protected by architecture rather than policy.

Principles

We ship the architecture, not promises.

  • Audio is destroyed before any note is written.
  • Every model runs locally or on-device.
  • The invariants are verifiable, and source access is available for review.
  • We do not train on your data. We never see it.
Who writes this

Kyle Nelson

Founder, Fazit · Berlin, Germany

hello@re-entry.ai

Kyle Nelson builds Fazit, a macOS notetaker that transcribes calls on the machine they happen on and never writes the audio to disk. He wrote the capture pipeline it describes: Core Audio process taps, a RAM-only ring buffer, on-device speech models, and a local model for the notes.

The writing here comes out of that work. Claims about what a tool does with your audio are checked against the vendor's own documentation and linked, because the interesting differences between these products are architectural and mostly undisclosed in marketing copy.

He is not a lawyer, an accountant, or a compliance professional. The posts that touch privilege, recordkeeping, consent law or GDPR cite the statute, the bar opinion or the court docket directly so you can read the source rather than take his summary of it. None of it is legal advice, and where a question is genuinely unsettled these pages say so.