Client & sales calls
Summarise contract terms, budgets and direct lines after reviewing the transcript and what the detector replaced.
Content-blind meeting recorder You talk.
You talk.
The booth makes a transcript.
The client call, the performance check-in, the clinical case conference — the meetings that most need a summary are exactly the ones you can't paste into a public AI notetaker. When its transcription model is available, cloak-meet transcribes audio in your browser. After review, the transcript is sent to the relay for an AI summary. Before sending, the browser replaces names, dates and identifiers its detector finds with realistic surrogates; values the detector misses may remain as written.
- Browser transcription when the model is available
- Detected values replaced before sending
- On-device receipt
or
Recording and transcription both run in this tab.
Five sentences — what the meeting was and what came out of it.
Record, upload or pick a sample to begin.
- from this site
- 0
- to the summariser
- 0
- anywhere else
- 0
Your summary will appear here.
The summary model receives the reviewed transcript. Values replaced during review are restored on this device when their surrogates return unchanged.
The meetings that can't go anywhere else
Performance check-ins
Turn performance check-ins into minutes after reviewing the transcript; values detected on-device are replaced before sending.
Clinical case conferences
Transcribe and summarise clinical case discussions; review the transcript and detected values before sending.
Board & leadership updates
Decisions and action items pulled out in any of the languages the on-device model already speaks — Latin, Cyrillic, CJK and more.
From voice to summary, in four moves
Record — in the booth
MediaRecorder captures your microphone (or you upload an audio file) straight into browser memory. This path does not upload the recording. This step sends nothing.
Transcribe — on-device
A Whisper speech model (transformers.js, same-origin, no HF hub) turns the audio into text entirely in your tab. The audio is transcribed in this tab; this path does not upload the recording.
Cloak — on-device
A WebAssembly engine plus a multilingual name model look for names, emails, phones and addresses in the transcript — across Latin, Cyrillic, Arabic and CJK — and replace detected values with realistic surrogates. Structured identifiers match by rule; free-form names rely on the on-device model, so an unusual one can slip past and remain as written. Review the request text before sending; the preview does not measure egress bytes. Detection sends no transcript text.
Summarise — re-inked
The reviewed transcript is sent through the CloakAPI gateway. The gateway adds a request integrity check. It does not verify that on-device checks ran. Values detected in the review are replaced by surrogates; some values can still pass both checks and remain as written. The summary returns with its surrogates for on-device re-identification. The receipt records detector layers and spans, but does not measure missed values.