Like Granola, but open source.
Silent meeting capture for macOS: auto-records your video calls (Zoom, Meet, Teams), transcribes with speaker attribution, writes structured notes + action items to plain files on disk. No bot joins the call. No virtual audio device. No backend account. Your LLM key, your prompt, your data — local by default. Any agent that can read a file can consume meeting context.
| Structured JSON output | Not just a pretty transcript UI — machine-readable data agents can consume |
| Plain files on disk | Claude Code, OpenClaw, or any agent reads ~/.quietclaw/meetings/ directly — no API, no setup |
| Open source | Apache 2.0, no vendor lock-in, extend it however you want |
| No meeting bot | Captures audio directly via macOS Core Audio Taps, invisible to other participants |
| Real-time transcription | Deepgram gives you streaming STT with built-in speaker diarization — bring your own API key, pay pennies per minute |
| Optional AI summarization | Claude extracts executive summaries, decisions, and action items — or skip it and let your agents process the raw transcript however they want |
Call detected (or started manually)
→ Audio capture runs in isolated utility process (no UI jank)
→ Mic + system audio sent as two mono streams to Deepgram
→ Speaker diarization separates participants on system channel
→ Calendar matcher names speakers from attendee list
Call ends
→ Transcript assembled with speaker attribution
→ Optional: Claude summarization → summary + action items
→ Files written to ~/.quietclaw/meetings/YYYY-MM-DD/{slug}/
→ Indexed in SQLite, your agents can now act on it
- macOS 13+ (Ventura or later — required for Core Audio Taps)
- Node.js 20+ and pnpm
- Xcode Command Line Tools (
xcode-select --install)
git clone https://github.com/laszlokorsos/quietclaw.git
cd quietclaw
pnpm install
pnpm run build:native # Build the native Core Audio addon
pnpm build # Build the app + package as DMGThe built .dmg is in dist/. Open it, drag QuietClaw to Applications, and launch. It runs as a menu bar app — look for the icon in your top bar.
On first launch, macOS will prompt for Screen Recording permission (required for audio taps).
Developer mode: If you want to hack on QuietClaw instead, run
pnpm devfor hot-reloading dev mode.
| Key | Required | Purpose |
|---|---|---|
| Deepgram | Yes | Real-time speech-to-text (~$0.0043/min) |
| Anthropic | No | AI summarization (Claude Haiku) |
Add keys via the Settings panel in the app, or set DEEPGRAM_API_KEY / ANTHROPIC_API_KEY environment variables for development.
Click Settings > Google Calendar > Connect Account to add one or more Google accounts. The app handles OAuth automatically — no GCP setup required from you.
QuietClaw writes plain files to disk — no API server, no MCP, no integration layer. Claude Code, OpenClaw, or any agent that can read files consumes meeting data directly.
Drop-in agent skill: the repo ships a SKILL.md that documents the on-disk layout, schema, and common query patterns as a reusable skill. Point Claude Code (or any agent framework with skill auto-discovery) at it and the agent learns how to answer "what were today's meetings", "what are my open action items", "what did I decide about X" without being told where anything lives.
# Today's meetings
ls ~/.quietclaw/meetings/$(date +%Y-%m-%d)/
# Read a transcript (JSON for parsing, Markdown for reading)
cat ~/.quietclaw/meetings/2026-04-05/weekly-standup-a1b2/transcript.json
cat ~/.quietclaw/meetings/2026-04-05/weekly-standup-a1b2/transcript.md
# Action items
cat ~/.quietclaw/meetings/2026-04-05/weekly-standup-a1b2/actions.json
# Daily index lists all meetings for a date
cat ~/.quietclaw/meetings/2026-04-05/index.mdEach meeting directory contains metadata.json, transcript.json, and optionally summary.json and actions.json. The Markdown files are the same data in human/LLM-readable format with YAML frontmatter. Full schema and richer query examples in SKILL.md.
Each meeting produces a directory under ~/.quietclaw/meetings/:
2026-04-04/
index.md # Daily index with links to all meetings
weekly-standup-a1b2/
metadata.json # Meeting metadata, speakers, calendar event
transcript.json # Timestamped, speaker-attributed segments
transcript.md # Human-readable transcript with YAML frontmatter
summary.json # Executive summary, topics, decisions (if summarized)
summary.md # Human-readable summary with YAML frontmatter
actions.json # Action items with assignees and priority
Files are plain JSON and Markdown — readable by any tool, diffable in git, and queryable with jq. See examples/ for complete sample output.
All markdown files include YAML frontmatter with structured metadata. Speaker names are rendered as wikilinks ([[Jordan]]). Point an Obsidian vault at ~/.quietclaw/meetings/ and your meetings become part of your knowledge graph — Dataview queries, backlinks, graph view all work out of the box.
- Silent audio capture via Core Audio Taps (mic + system audio, no virtual device)
- Auto-detection of Google Meet, Zoom, and Teams calls via window title and mic activity
- Real-time transcription via Deepgram with multi-speaker diarization
- WebRTC AEC3 echo cancellation on the mic path — the same echo canceller Chrome's
getUserMediaships, which is what Google Meet uses in a browser. Good baseline, not state-of-the-art (commercial stacks like Teams and Zoom layer a trained neural residual suppressor on top — see ARCHITECTURE.md for the honest quality comparison and how to swap in DeepFilterNet 3 if you want to push further). - Speaker identification — mic audio is always "you"; system audio speakers separated by diarization; 2-person calls fully named via calendar; 3+ person calls support manual speaker mapping
- Google Calendar integration — multi-account OAuth, automatic event matching, attendee extraction
- AI summarization — Claude turns the transcript into topic-organised meeting notes with action items. Highly recommended but optional; transcripts still save without it.
- Structured output — JSON + Markdown files per meeting, indexed in SQLite
- Crash recovery — orphaned recordings automatically recovered on next launch
- Platform join buttons — upcoming meetings show clickable Google Meet / Zoom / Teams buttons
- Obsidian-compatible — YAML frontmatter, wikilinks, daily index files
QuietClaw uses a TOML config file at ~/.quietclaw/config.toml. Key settings:
[general]
data_dir = "~/.quietclaw/meetings" # Where meetings are stored
[stt]
provider = "deepgram" # STT provider (deepgram is the default)
[stt.deepgram]
model = "nova-3" # Deepgram model
language = "en"
diarize = true # Multi-speaker separation
[summarization]
enabled = true # Set false to skip summarization
provider = "anthropic"
model = "claude-haiku-4-5-20251001"
See resources/default_config.toml for all options.
- Contacts database — persistent people across meetings, email-to-name mapping, so agents can query "all meetings with Alice"
- Speaker consistency — autocomplete from past names, stable wikilinks across meetings
- Cross-meeting queries — "all meetings with person X", "open actions for team Y"
- Voice fingerprint learning — automatic speaker recognition that improves from manual corrections
- Meeting type templates — different summarization prompts for standups, 1:1s, all-hands
- Microsoft/Outlook calendar — enterprise calendar support via Graph API
- Windows support — WASAPI loopback capture (architecture is already abstracted behind
AudioCaptureProvider) - Webhooks — push meeting data to n8n, Zapier, or custom endpoints when processing completes
- Multi-language — transcription and summarization in non-English languages
pnpm dev # Dev mode with hot reload
pnpm test # Run vitest
pnpm typecheck # TypeScript strict mode check
pnpm build # Full production build (vite + electron-builder)See CLAUDE.md for the full development guide — architecture, coding standards, how to add new providers.
| Layer | Technology |
|---|---|
| App framework | Electron (TypeScript) |
| Audio capture | macOS Core Audio Taps (native N-API addon) |
| STT | Deepgram (real-time WebSocket streaming) |
| Summarization | Anthropic Claude (Haiku default) |
| Calendar | Google Calendar API (OAuth) |
| Database | better-sqlite3 |
| UI | React 19 + Tailwind CSS |
| Config | TOML (~/.quietclaw/config.toml) |
| Secrets | Electron safeStorage (OS-level encryption) |
Contributions are welcome! See CONTRIBUTING.md for setup instructions and the PR process.
For security vulnerabilities, see SECURITY.md — do not open a public issue.
QuietClaw records audio from your meetings. Recording laws vary by jurisdiction — some require only one party's consent, others require all participants. It is your responsibility to understand and comply with the recording laws that apply to you.
App icon and tray glyph are based on "Crab claw" by Lorc, licensed under CC BY 3.0. Colour recast and compositing by the QuietClaw authors.