Folio

Local-first meeting transcription for macOS. Captures system audio + microphone independently, transcribes on-device, writes per-meeting markdown to your vault. Audio never leaves your machine. Website: folio.chele.bi.

Status: 1.0.0. Install via Homebrew or the notarized DMG, or build from source.

What it does

• Sits in the menu bar. Watches your calendar and audio devices via EventKit (no Google OAuth, no Microsoft Graph).
• When a meeting starts, records both system audio and your microphone as separate streams via cpal + ScreenCaptureKit.
• Transcribes locally by default. The bundled Whisper backend (whisper.cpp via whisper-rs, Metal-accelerated on Apple Silicon) is the primary path; OpenAI's Whisper API is an opt-in fallback for users who want a faster cloud transcription on long meetings.
• Diarizes the system-audio track on-device (Speaker 1, 2, 3…) with pyannote segmentation + a speaker-embedding model; your microphone is always labelled You. No cloud.
• Writes a markdown file per meeting to your chosen vault path. Frontmatter includes attendees, duration, model, source.
• Ships a local MCP server (folio-mcp). Any MCP-aware tool — Claude Desktop, Cursor, Claude Code — gets read-only access to your transcripts, tasks, and memories over stdio. No cloud, no proxy (Settings → Connectors).
• Audio never leaves your machine on the default path. Privacy Mode (Settings → Privacy) physically blocks every outbound HTTP call except localhost so the app keeps working end-to-end with Wi-Fi off.

Privacy & consent

No telemetry, no analytics, no crash reporting — enforced in CI. Audio, transcripts, and notes stay on your machine on the default path; the only network calls are the opt-in cloud-AI and webhook paths plus the one-time model download, all of which Privacy Mode can block. Full details, including data retention, are in docs/PRIVACY.md.

Recording a conversation can be illegal without the other participants' consent, and the rules vary by US state and by country (many require all-party consent). Folio gives you the tool; obtaining consent is your responsibility. Tell people before you record. See docs/PRIVACY.md.

Install

Requires macOS 13 (Ventura) or later, on Apple Silicon or Intel. The first time you use local transcription or diarization, Folio downloads the model weights (a few hundred MB) from Hugging Face and the sherpa-onnx GitHub releases.

Homebrew (recommended)

brew tap woosal1337/folio https://github.com/woosal1337/folio
brew install --cask folio

brew upgrade --cask folio tracks new releases.

Direct download

Grab the latest Apple Silicon .dmg from the Releases page, open it, and drag Folio to Applications:

• Apple Silicon — Folio_<version>_aarch64.dmg

Releases are code-signed with a Developer ID and notarized by Apple, so they open without a Gatekeeper prompt. (Intel builds aren't published yet — build from source on an Intel Mac if you need one.)

Build from source

See Development.

Repository layout

folio/
├── Cargo.toml                # workspace root
├── rust-toolchain.toml       # Rust 1.88, both Apple targets
├── crates/
│   ├── folio-core/          # audio capture, storage, transcription, diarization
│   ├── folio-cli/           # CLI test harness
│   └── folio-mcp/           # local MCP stdio server (notes, tasks, memories)
├── src-tauri/                # Tauri 2 desktop binary
├── src/                      # React 18 + TypeScript + Tailwind frontend
├── Casks/                    # Homebrew cask (this repo doubles as its own tap)
├── docs/                     # repo-local docs (see ARCHITECTURE.md)
└── .github/workflows/        # CI + release

For the on-disk module map, the IPC contract, and the rationale behind the key design decisions (why mic and system audio are captured as separate streams, why the default transcription path is local Whisper rather than a cloud API), see docs/ARCHITECTURE.md.

Stack: Rust core (~70%) for audio + storage + transcription, Tauri 2 wrapping a React + TypeScript + Tailwind frontend. cpal for mic capture, ScreenCaptureKit for system audio, rubato for resampling, hound for WAV. TypeScript bindings for IPC types are generated from Rust via ts-rs.

Development

Requirements:

• macOS 13 (Ventura)+ on Apple Silicon (Intel Macs build fine; Apple Silicon is the perf target)
• Rust 1.88 via rustup (the toolchain is pinned in rust-toolchain.toml)
• Bun 1.3+ (the only JS package manager + runtime this repo uses)
• Xcode command-line tools: xcode-select --install

git clone git@github.com:woosal1337/folio.git
cd folio
bun install
pre-commit install
pre-commit install --hook-type commit-msg
pre-commit install --hook-type pre-push

Run the desktop app

bun tauri dev

First launch compiles the Rust workspace (~30 s on a warm cache). Subsequent launches reuse the cache.

Run the CLI test harness

# List input devices
cargo run -p folio-cli --release -- devices

# Record from the default device for 60 seconds
cargo run -p folio-cli --release -- record --seconds 60

# Record from a specific device
cargo run -p folio-cli --release -- record --seconds 60 --mic-device "MacBook Pro Microphone"

Output: ./recordings/<timestamp>/mic.wav (mono 16-bit PCM at the device's native rate).

On first run macOS prompts for microphone permission (and screen recording permission for system audio capture).

Local checks (mirror CI)

# Rust
cargo fmt --all
cargo clippy --workspace --all-targets -- -D warnings
cargo test --workspace --lib --bins
cargo deny check

# Frontend (bun runs the npm scripts defined in package.json)
bun run typecheck
bun run lint
bun run format:check
bun run test

pre-commit runs the relevant subset on every commit; the full suite also runs in .github/workflows/ci.yml.

Contributing

See CONTRIBUTING.md for setup and the workflow. The authoritative code-style contract is docs/CODE_STYLE.md. Security issues go through SECURITY.md.

License

Apache-2.0. See LICENSE. Third-party attributions in NOTICE.