Skip to content

README.md

Latest commit

 

History

History
206 lines (162 loc) · 8.43 KB

README.md

File metadata and controls

206 lines (162 loc) · 8.43 KB

Sidekick2000

A macOS desktop app that records meetings, transcribes them, and produces structured notes — with action items pushed directly as GitHub issues and notes committed to a git repository.

Built with Tauri + Svelte 5 on the frontend, Rust on the backend.

What it does

  1. Records two audio streams in parallel — local mic and remote source (system audio via BlackHole) — with live dual VU meters
  2. Transcribes live both streams in real time using local whisper.cpp with Metal GPU acceleration — no audio leaves your machine. VAD (Silero) detects speech boundaries and feeds chunks to Whisper as you speak, so the transcript appears during the meeting.
  3. Merges the two transcripts into a single time-sorted conversation, each segment labelled with the correct speaker name (one stream = one speaker, no diarization needed)
  4. Summarizes with Claude Sonnet (Anthropic) or any Together.ai chat model, using a context you define (meeting type, participants, domain vocabulary)
  5. Exports structured Markdown notes (YYYY-MM-DD_HHmm_Context.md) to a configurable folder
  6. Commits the notes to a git repository automatically (optional)
  7. Creates GitHub issues for every action item extracted from the notes (optional)

Privacy: All audio is processed locally on-device. Only the final text transcript is sent to the cloud for summarization (Anthropic or Together.ai). A Groq Whisper cloud fallback is available in settings if preferred.

Output format

Each meeting produces two files:

Meetings/
  2026-03-20_1430_welqin.md          ← structured notes
  2026-03-20_1430_welqin_transcript.md ← raw transcript with named speakers

The notes follow a consistent structure:

## Participants
## Summary
## Key Discussion Points
## Decisions Made
## Action Items
- [ ] **David**: Review the API design document
- [ ] **Yannick**: Set up CI pipeline for staging

Action items are automatically created as GitHub issues with the meeting-action label.

Setup

Requirements

  • macOS Apple Silicon (M1 or later — Metal GPU required for local Whisper)
  • BlackHole 2ch — virtual audio driver for capturing system audio (see setup guide)
  • cmake — required to build whisper.cpp (brew install cmake)
  • One of (for summarization):
  • gh CLI installed and authenticated — for GitHub issues (optional)
  • git — for committing notes (optional)
  • Groq API key — only if using cloud transcription fallback (optional)

The Whisper model (large-v3-turbo, quantized q5_0, ~550 MB) is downloaded automatically on first launch to ~/.sidekick2000/models/.

Install dependencies

npm install

Configure

Launch the app and click the gear icon in the top-right corner. All settings are stored in ~/.sidekick2000/settings.json.

Tab What to configure
API Keys Transcription mode (Local Whisper or Groq); summarization provider (Claude or Together.ai) and the corresponding API key/model
Devices Local mic device + your speaker name; remote source device + remote speaker name
Repository Working folder (git root), meetings subfolder, GitHub repo (owner/repo), default language, pipeline step toggles
Contexts Meeting context templates — instructions that shape how the AI summarizes each meeting type
Speakers Default meeting attendees pre-loaded at startup (for AI context)

API keys set in Settings take priority over environment variables. You can still use a .env file as fallback.

Dev mode

npm run tauri dev

Build

npm run tauri build

Contexts

Contexts are the core of Sidekick2000's flexibility. Each context is a Markdown document that gives Claude background knowledge about a meeting type: who the participants are, domain vocabulary, and how to structure the notes.

Examples of contexts you might create:

  • General — neutral instructions, works for any meeting
  • Product review — focus on decisions, feature requests, backlog items
  • Client call — highlight commitments, risks, next steps
  • Training session — track exercises, Q&A, shortcuts mentioned

Contexts are managed entirely in the Settings UI (no external files needed).

Pipeline

Record local mic  ─────────────────┐   (parallel, ring buffers)
Record remote source ───────────────┤
                                    │
                    ┌───────────────┴───────────────┐
                    ▼                               ▼
           Worker thread                   Worker thread
           (drain every 200ms)             (drain every 200ms)
                    │                               │
                    ▼                               ▼
           VAD (Silero)                    VAD (Silero)
           silence ≥ 300ms → flush         silence ≥ 300ms → flush
                    │                               │
                    ▼                               ▼
           Whisper (Metal)                 Whisper (Metal)
           whisper.cpp local               whisper.cpp local
                    │                               │
                    ▼                               ▼
           emit "live-segment"             emit "live-segment"
           → frontend display              → frontend display
                    └───────────────┬───────────────┘
                                    │
                                    ▼
                    Merge — sort by timestamp, speakers already known
                                    │
                                    ▼
                    Summarize  (Claude Sonnet or Together.ai — skipped if disabled)
                                    │
                                    ▼
                    Export  YYYY-MM-DD_HHmm_Context.md
                                    │
                                    ▼
                    Git commit  (if enabled and working folder configured)
                                    │
                                    ▼
                    Create GitHub issues  (if enabled and repo configured)

Settings file

~/.sidekick2000/settings.json — created automatically on first save.

{
  "transcription_mode": "LocalWhisper",
  "groq_api_key": "",
  "anthropic_api_key": "sk-ant-...",
  "together_ai_api_key": "",
  "summarization_provider": "claude",
  "together_ai_model": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
  "default_input_device": "MacBook Pro Microphone",
  "local_speaker_name": "David",
  "remote_device": "BlackHole 2ch",
  "remote_speaker_name": "Remote",
  "working_folder": "/Users/you/my-repo",
  "github_repo": "owner/repo",
  "meetings_subfolder": "Meetings",
  "default_language": "fr",
  "enable_summary": true,
  "enable_git_commit": true,
  "enable_github_issues": true,
  "default_speakers": [
    { "name": "Alice", "organization": "Acme" }
  ],
  "contexts": [
    {
      "id": "general",
      "label": "General",
      "content": "Be factual. Group by theme. Use professional tone."
    }
  ]
}

Tech stack

Layer Technology
UI Svelte 5, Tailwind CSS 4
Desktop shell Tauri 2
Backend Rust (async with Tokio)
Transcription Local whisper.cpp via whisper-rs (large-v3-turbo q5_0, Metal GPU). Groq Whisper API as optional cloud fallback.
VAD Silero VAD via voice_activity_detector (ONNX) — detects speech/silence for live chunking
Summarization Anthropic Claude Sonnet or Together.ai (configurable)
Speaker identification Device-based — each stream has a pre-assigned speaker name
Audio capture CPAL (two simultaneous input streams, ring buffers, shared t=0 origin)
GitHub integration gh CLI