Minutiae

The privacy-first AI notetaker for productive meetings. — minutiae.dev

Local-first meeting notepad for macOS. Minutiae captures your mic and the other side of the call without a meeting bot, shows a live two-channel transcript beside your notes, and, when the meeting ends, turns both into a Markdown note in a folder you choose. Transcription and enhancement run on the Mac in front of you. Nothing is sent anywhere.

Requires an Apple Silicon Mac on macOS 14.4 or later.

What it does

• No bot joins the call. System audio is captured with a Core Audio process tap and the mic with AVAudioEngine, so it works with Zoom, Meet, FaceTime or a YouTube tab alike.
• Live transcript, two speakers. Every line carries a Me / Them badge and a timestamp. On-device ASR (NVIDIA Parakeet TDT v3 via FluidAudio, on the Neural Engine) transcribes utterance by utterance as people speak.
• Silence produces nothing. An RMS gate runs before the model, so a pause is a pause, not a hallucinated sentence. There is a test for it.
• Echo suppression. On speakers the far end reaches the mic too; Minutiae uses the system-audio tap as a reference so it is not transcribed twice. It never clips your own voice.
• Notes while you listen. A scratchpad is saved with the session.
• Enhance, on-device. A local language model (Qwen3.5-4B via MLX, one 2.9 GB download) merges transcript and notes into a finished note with YAML frontmatter.
• Plain files. Sessions are JSON, Opus audio and Markdown on disk. Notes land in any folder, so an Obsidian vault opens them as-is.
• Small. Idle footprint under 250 MB, capture starts in under a second, no telemetry, no network calls except model downloads.

Install

1. Download the latest Minutiae.dmg from Releases and drag Minutiae to Applications.
2. On first launch macOS asks for microphone and system-audio recording permission. Both are required to capture a call.
3. The app downloads the ASR model (about 1.5 GB, once) and compiles it for the Neural Engine. Recording is available as soon as that finishes.
4. Pick a notes folder in the sidebar (Vault). The language model for Enhance is downloaded only when you first ask for it.

Sessions live in ~/Library/Application Support/Minutiae/sessions/; the format is documented in docs/session-format.md.

Build from source

Requirements: Apple Silicon, macOS 14.4+, Xcode 16.4+ (Swift 6.1), Rust stable, Node 22, pnpm 10.

pnpm install
pnpm dev                      # builds the Swift engine sidecar, then runs `tauri dev`

# Optional, for the Enhance feature in a dev build (slow MLX build, run once):
scripts/build-llm-sidecar.sh release

Tests:

swift test --package-path engine                       # audio engine
cargo test --manifest-path app/src-tauri/Cargo.toml    # Rust core
cargo test --manifest-path engine-windows/Cargo.toml   # Windows sidecar (portable parts)
pnpm --filter minutiae check                           # svelte-check

See CONTRIBUTING.md for the development workflow and docs/signed-build.md for producing a signed, notarized release.

How it is put together

┌──────────────────────────────┐   NDJSON over stdio   ┌──────────────────────────────┐
│ Tauri 2 app                  │ ◄───────────────────► │ minutiae-engine  (Swift)     │
│ Rust core + Svelte 5 UI      │ sidecar-ipc-v1        │ process tap + mic capture    │
│ sessions, settings, recents, │                       │ echo suppression             │
│ prompt assembly, vault notes │                       │ FluidAudio ASR on the ANE    │
│                              │   NDJSON over stdio   ├──────────────────────────────┤
│                              │ ◄───────────────────► │ minutiae-llm  (Swift, MLX)   │
│                              │ llm-ipc-v1            │ Qwen3.5-4B completion engine │
└──────────────────────────────┘                       └──────────────────────────────┘

Audio never crosses the process boundary: the engine writes the audio files itself and only control messages and transcript events travel over stdio. docs/architecture.md walks through a session end to end and lists the invariants the design depends on.

Directory	Contents
app/	Tauri 2 application: Rust core (app/src-tauri) and Svelte 5 UI (app/src)
engine/	Swift audio and ASR sidecar (SwiftPM package)
llm-engine/	Swift MLX sidecar for on-device enhancement
engine-windows/	Rust sidecar for the Windows port (phase 1: IPC and device enumeration)
docs/	Protocol specs, session format, architecture, milestones, signed builds
scripts/	Build, packaging, signing and icon scripts

Docs

• docs/architecture.md, how the pieces fit and why
• docs/protocol/sidecar-ipc-v1.md, the engine wire protocol
• docs/protocol/llm-ipc-v1.md, the LLM sidecar wire protocol
• docs/session-format.md, what a session folder contains
• docs/milestones.md, where the project is and where it is going
• docs/windows-port.md, the Windows design

Open core

The public build has no account, no sync and no cloud. A separate, closed cloud tier (sync of session text and cloud enrichment) exists as a client module that is not in this repository; the saas Cargo feature and the VITE_SAAS build flag are the hooks it plugs into. Without that module the feature does not build, and the code you see here is the whole product for everyone else.

Status

Milestones 1 and 2 (capture, live transcript, scratchpad, on-device enhance) are shipped. Signed and notarized builds are scripted. The Windows sidecar has its IPC seam in place but no capture yet. See docs/milestones.md.

License

MIT. See LICENSE. Models downloaded at runtime and third-party libraries have their own licenses; see THIRD_PARTY.md.