Recall
Product & Desktop Engineer
About the Project
Recall captures your microphone and the other side of any desktop call — Zoom, Meet, Teams, WhatsApp Desktop, browser calls — streams it to Deepgram for a live transcript, and summarises the call with an LLM when you stop. It is deliberately local-first: audio, transcripts and summaries live on your machine in SQLite, and the only outbound traffic in the entire app is audio to Deepgram and transcript text to OpenAI. No backend, no telemetry, no cloud sync. Built as an Electron app with packaged installers for Windows and macOS. Because there is no server, the architecture is a process split rather than a service split, and the boundary between those processes is where all the interesting decisions live.
What It's Made Of
Renderer — capture & UI
React · AudioWorklet · MediaRecorder
The window, and deliberately also the audio capture layer: MediaStream and getUserMedia do not exist in Electron's main process, so device access has to live here. It opens both devices, merges mic into channel 0 and system audio into channel 1 — an ordering the transcriber relies on to attribute speech without guessing — and streams 16 kHz PCM to main while a parallel MediaRecorder writes the archive. It reaches main only through a typed bridge; no renderer code touches IPC directly.
Main process — transcription & summarisation
Electron main · Deepgram WebSocket · OpenAI
Holds everything privileged: the Deepgram streaming socket producing the live diarized transcript, the post-call summarisation pass that extracts action items and decisions, API keys encrypted through the OS keychain rather than left in a config file, and a privileged media protocol for playback that verifies a requested file really belongs to that call before serving bytes, with byte-range support so the player can seek.
Local storage
SQLite · inlined migrations · Opus-in-WebM
Calls, transcripts, summaries and settings in SQLite alongside the audio files on disk. Migrations are compiled into the bundle rather than read from the filesystem, because directory reads break inside a packaged app, and they run idempotently against a migrations table. The schema layer is kept Electron-free so it is testable under plain Node.
Shared logic & packaging
Vitest · electron-builder · Windows + macOS
Pure logic — formatting, Markdown export — lives in a shared module usable from both processes and testable without Electron, under Vitest. Packaging produces signed installers for both platforms, including a native-module rebuild step against Electron's ABI, with a separate Node-ABI copy of the database driver so tests can run outside Electron.
Key Highlights
- Captures mic and system audio together — WASAPI loopback on Windows, and a guided BlackHole virtual-device setup on macOS, where the OS gives no direct route to system audio
- Streams 16 kHz PCM from an AudioWorklet over IPC to a Deepgram WebSocket for a live diarized transcript, while a parallel MediaRecorder writes the audio archive
- Generates a structured post-call summary with action items and decisions via OpenAI, stored alongside the transcript in SQLite
- Encrypts API keys through the OS keychain (Windows Credential Manager, macOS Keychain) rather than leaving them in a config file
- Serves playback through a privileged protocol handler that verifies the requested file belongs to that call and implements byte ranges itself, since without them the player cannot seek
- Added optional screen recording that raises the consent bar deliberately — the per-call checkbox rewords itself when video is on, and video never leaves the machine or reaches Deepgram
Technical Challenges
The architectural constraint is that MediaStream APIs do not exist in Electron's main process, so audio capture has to live in the renderer while transcription, storage, summarisation and key handling stay in main — and the renderer never touches IPC directly, only a typed bridge from the preload script. The decision I care about more is that consent is a feature, not a disclaimer: many jurisdictions require every participant to agree, so Recall never records on its own, shows a red indicator and tray icon the entire time, and degrades to microphone-only with an explicit message rather than silently failing when system audio is unavailable.