RecordAI
Record a call or negotiation entirely on your Mac: two channels “you / the other side”, local GPU transcription, live translation by a local model and summaries through your own n8n.
Request access →- Format
- Desktop app for macOSmacOS 13+, Apple Silicon; one .app with no dependencies
- Status
- Pilot
- Stack
- Tauri 2 · Rust · whisper.cpp + llama.cpp (Metal) · ScreenCaptureKit · n8n
- Direction
- Speech & Audio
About the product
RecordAI is a Mac app that records a call or an in-person meeting and turns it into text without sending audio anywhere. You press one button (or the app offers to start when something takes the microphone), and when you stop you get an MP3, a transcript labeled “you / the other side”, optionally live translation as the conversation goes, and a summary with agreements and tasks. Everything sits in a plain folder under Documents and opens without the app.
It is for people who have more calls than time to retell them: managers and project leads, sales and account teams, HR on interviews, consultants and lawyers, teams negotiating in a foreign language. Instead of a cloud “AI notebook” with a subscription and someone else's servers — one .app installed by drag and drop that works offline; instead of hand-written notes — a ready summary in your n8n, from where it goes to Notion, Slack or your CRM.
Two sources
Microphone and system audio — two channels of one file
60 s chunks
Audio goes to disk at once: a crash loses at most a minute
whisper on Metal
Each channel recognized separately, locally
Text and translation
“You / other side” transcript, translation by a local LLM
Summary in n8n
Only text — to your webhook, markdown comes back
The key difference is two-channel recording instead of diarization: “who spoke” is known exactly because you and the other side are recorded into different channels, and no ML model guesses voices. The second is where the privacy line runs — at the audio: recording and translation are computed on the Mac, only text goes out, and only to your own n8n. The app has no server, account or subscription of its own; at startup it can ask your n8n for permission to run — that is how a company manages installs by a list of machines.
What it looks like
One window with four screens plus a tray icon. Frames are from a demo recording; click to enlarge.
Record
Recording card: summary
Recording card: transcript
Translation
Settings
Arrows, dots or drag with the mouse; in the enlarged view — ← → and Esc.
Who it is for
Teams with frequent calls
A “you / other side” transcript right after stopping, a summary with tasks a minute later — one click to your wiki or tracker.
Negotiations in another language
Live translation of the other side's lines in a separate feed — by a local model, no cloud translator and no network latency.
Companies with privacy requirements
Audio stays on the Mac, translation runs on the Mac, text goes only to your own n8n. No intermediary service.
How it works
- 1
Start recording
A button in the window, a tray item, or the app's own offer when a call takes the microphone. Zoom, Meet, Telegram, a browser — the source does not matter. Pause, and the recording continues from the same spot.
- 2
Talk as usual
While the meeting goes on, the transcript grows in the background, and the Translation screen shows the other side's lines in Russian or English — the window closes at a pause in speech, at a phrase boundary.
- 3
Stop
Chunks are assembled into an MP3, transcription catches up, the recording card opens with audio, transcript and timestamps. Nothing to download or upload.
- 4
Get a summary
The “Summarize” button sends the text to your n8n; outcomes, agreements and a task table come back. Copy as Markdown, HTML or BBCode and pass it on — to a wiki, an email, a tracker.
What the product does
Two-channel recording
Microphone (cpal) and system audio (ScreenCaptureKit) are recorded separately into one stereo MP3: left channel — you, right — the other side. Echo cancellation in Rust (NLMS + residual suppression) removes the other side leaking from the speakers into the microphone — otherwise their lines would land in the transcript twice.
Pause, tray, auto-record
Start, pause and stop from the window or the tray icon; the window can be closed. Pause does not break the file, the timer counts only recorded time. The app watches the microphone: when a call takes it, it offers to start recording or starts on its own. Launch at login, a two-hour safety stop.
A recording you cannot lose
Audio is flushed to disk in 60-second chunks: memory does not depend on meeting length, and kill -9 loses at most the last minute. An unfinished recording is picked up on the next launch — the MP3 and transcript are assembled from the chunks already on disk.
Local transcription
whisper.cpp on Metal: each channel is recognized separately, VAD drops silence, the language is auto-detected. Three models — from fast base to large-v3-turbo, downloaded from the app with SHA-256 verification. A finished recording can be re-transcribed with another model.
Live translation by a local model
The translation window closes at a pause in speech (3–12 s, adjustable) — a line ends at a phrase boundary, not mid-word. Qwen2.5-Instruct via llama.cpp on the same GPU, no network; the ~2 GB model is loaded only while translating and unloaded after the recording. Russian or English; only the other side, only you, or both.
Summaries via your n8n
The transcript goes as a POST with Basic Auth to your webhook and markdown comes back: headings, lists, task tables. The prompt, the model and the onward route are on your side. Copy as Markdown, HTML or BBCode with one button.
Recording cards
A list of meetings with duration and “transcript” / “summary” badges. The card has a player, rename, delete, the channel-highlighted transcript and the summary. One folder per recording under Documents: audio, transcript, summary, metadata — no database.
Managing installs
At startup the app asks your n8n for permission to run, sending its version and a device identifier (a hash of the hardware UUID that survives reinstalls). Keep a list of Macs, show messages or block an install. Without network the app still works — only an explicit refusal blocks it.
Privacy by construction
Audio never leaves the computer: recognition and translation run on the Mac, only text goes out — and only to your n8n. No intermediary service, subscription or account. One .app without Python, ffmpeg or Homebrew, works offline, fonts and models stored locally.
More products in this direction
Want to see the product on your data?
We will show a demo, walk through your scenario and suggest a deployment format.