Echoic

Native macOS meeting companion that captures audio from any meeting app, transcribes it locally, identifies speakers, and generates structured summaries. No bots, no plugins, no cloud audio upload.

How it works

Echoic sits in your menu bar. Click record before a call and it captures system audio via ScreenCaptureKit — works with Zoom, Google Meet, Teams, or any browser-based call. Transcription and speaker diarization run on-device using CoreML (WhisperKit + SpeakerKit). Only transcript text is sent to the Anthropic API for summarization. Audio never leaves your machine.

Features

• Universal capture — records system audio at the OS level, no per-app integration needed
• Real-time transcription — WhisperKit (CoreML) transcribes as you talk, searchable during recording
• Speaker diarization — identifies 2-6 speakers, optional mic input for "You" labeling
• Structured summaries — decisions, action items, and open questions via Anthropic API
• Full-text search — FTS5-powered search across all meeting transcripts
• Local-first — audio stored as standard AAC, transcripts in SQLite, all in ~/Library/Application Support/Echoic/
• Crash-safe — audio flushed to disk every 30 seconds

Requirements

• macOS 13 (Ventura) or later
• Apple Silicon recommended (Intel supported with batch transcription fallback)
• Screen Recording permission (for system audio capture)
• Anthropic API key (for summarization)

Building

# Build
xcodebuild -scheme Echoic -configuration Debug build

# Run tests
xcodebuild -scheme Echoic -configuration Debug test

Setup

1. Build and run the app
2. Grant Screen Recording permission when prompted
3. Open Settings from the menu bar icon and enter your Anthropic API key
4. Click Start Recording before your next call

Architecture

Echoic/
├── App/            # Entry point, AppDelegate, menu bar
├── Audio/          # ScreenCaptureKit capture, AVAudioEngine mic, ring buffer, AAC encoding
├── Transcription/  # WhisperKit integration, chunked inference, VAD, dedup
├── Diarization/    # SpeakerKit integration, speaker merge, local speaker ID
├── Summarization/  # Anthropic API client, prompt templates, retry logic
├── Storage/        # GRDB.swift database, migrations, FTS5 search
├── Models/         # Data models (Meeting, TranscriptSegment, Summary, Speaker)
├── Views/          # SwiftUI — Library, MeetingDetail, Settings, Onboarding
├── Services/       # Keychain, model download, audio playback
└── Utilities/      # ULID generation, date formatting

Single-process architecture — no sidecars, no XPC services. WhisperKit and SpeakerKit run in-process via CoreML.

Privacy

• Audio is captured and processed entirely on-device
• Only transcript text is sent to the Anthropic API for summarization
• API keys stored in macOS Keychain
• All meeting data lives in user-accessible standard formats (SQLite, AAC, JSON)

License

This project is licensed under the GNU Affero General Public License v3.0.