Harknotes

Local-first meeting notes for Linux (macOS planned). A privacy-focused Granola alternative: no meeting bot, no cloud account, audio never leaves your machine.

• Silent capture — records your mic and system audio (the other side of the call) directly from PipeWire/PulseAudio. No bot joins the meeting.
• Live local transcription — whisper.cpp transcribes ~45s chunks during the recording, one bounded process at a time, so the transcript is ready moments after the meeting ends — even on weak hardware.
• Crash-proof — every chunk is persisted and resumable; the in-progress buffer is snapshotted every 5s. A hard crash loses at most ~5 seconds of audio.
• Local or BYOK summaries — structured summaries (topics, action items, decisions, open questions) via any OpenAI-compatible endpoint: Ollama / LM Studio / llama.cpp locally, or your own OpenRouter/Groq/OpenAI key. Only transcript text is ever sent — and only if you ask for a summary.
• Your data — everything lives in a local SQLite DB; one-click markdown export.

Stack

Nx + pnpm monorepo (scaffolded by the forge):

• apps/web — React 19 + Vite + TanStack Router/Query + Tailwind 4 (the UI)
• apps/desktop — Tauri v2. All native concerns in Rust: audio capture (parec monitor + cpal mic), silence-boundary chunker, whisper sidecar worker, SQLite (rusqlite), markdown export, tray
• packages/* — shared config/types

Development

System deps (Fedora):

sudo dnf install -y webkit2gtk4.1-devel openssl-devel libappindicator-gtk3-devel \
  librsvg2-devel gtk3-devel alsa-lib-devel gcc-c++ cmake

Build the whisper.cpp sidecar (once per machine — it's gitignored):

bash apps/desktop/scripts/build-whisper.sh

Run:

pnpm install
cd apps/desktop && pnpm tauri dev

Test:

cd apps/desktop/src-tauri && cargo test   # Rust: chunker, resampler, db, export
cd apps/web && pnpm test                  # TS: summarizer parsing/coercion

Package (deb/rpm/AppImage):

cd apps/desktop && pnpm tauri build

Architecture notes

• Dual-stream: mic → "User" segments, system-audio monitor → "Meeting" segments, merged by absolute timestamp. Whisper models auto-download from Hugging Face on first use (tiny/base/small/medium/large-v3-turbo).
• Silence gate: chunks below an RMS floor are never transcribed — this is what kills whisper's "thank you for watching" hallucinations on dead air.
• One whisper at a time: the worker is a single queue with nice -n 10, a thread cap, a hard per-chunk timeout, and one retry. Failed chunks are retried on next launch.

Status

Linux is fully working (capture → live transcription → notes → summaries → export); deb/rpm bundles build at ~6MB. Known gaps: AppImage bundling fails under linuxdeploy on Fedora (deb/rpm unaffected), and macOS capture (ScreenCaptureKit) is the next platform target.

License

Harknotes is free software, licensed under the GNU AGPL-3.0. Forks and derivatives — including hosted versions — must remain open source.