Nextcloud Talk AI Notetaker

This bot will automatically summarize any Nextcloud Talk calls it's a participant of, using Google Gemini AI. Notes are saved to Nextcloud or locally, and optionally emailed to participants.

How It Works

Poll loop (10s)
  |
  +-- monitor.py    -- Talk API: detect active calls
  +-- recorder.py   -- Playwright + WebRTC: join call, record audio
  +-- transcriber.py -- Gemini AI: transcribe + summarize (chunked pipeline)
  +-- storage.py    -- WebDAV: upload notes to Nextcloud
  +-- mailer.py     -- SMTP: email notes to participants

The bot joins calls as a regular participant, captures remote audio through in-browser WebRTC interception, and produces structured meeting notes with topics, action items, and decisions.

Quick Start

git clone https://github.com/your-user/nextcloud-talk-notetaker.git
cd nextcloud-talk-notetaker
./setup.sh

The setup wizard will guide you through:

1. Checking Docker is installed
2. Connecting to your Nextcloud
3. Setting up a Gemini API key (free)
4. Choosing where to save notes
5. Optionally configuring email notifications

Prerequisites

• Linux with Docker (Ubuntu, Debian, Fedora, Arch, or any distro)
• Nextcloud with the Talk app installed
• A dedicated bot user account in Nextcloud
• Gemini API key (free at aistudio.google.com/apikey)

Docker handles all internal dependencies (Python, ffmpeg, Chromium).

Nextcloud User Setup

1. Create a bot user

In your Nextcloud admin panel, create a new user (e.g., ai-notetaker). This account does not need admin privileges.

2. Add bot to Talk rooms

Add the bot user as a participant in any Talk room you want it to monitor. The bot will automatically join active calls and start recording.

3. Set up notes folder (if using Nextcloud storage)

Create a folder in your Nextcloud (e.g., meeting-notes) and share it with the bot user. The bot saves notes into this shared folder.

4. Set up subadmin for email lookup (optional)

For the bot to look up participant email addresses, it needs subadmin status for user groups. This is not the same as admin — it only grants read access to user info within those groups.

In the Nextcloud admin panel:

1. Go to Users > Groups
2. For each group whose members should receive email notifications, add the bot user as a subadmin (not a regular member)

5. Notes filename

Notes are saved as YYYY-MM-DD-HHMM-firstname-firstname.md, with participant first names in alphabetical order. Set FILENAME_LAST_USER to always place a specific user's name last.

Configuration

All configuration is via environment variables in .env:

Variable	Required	Default	Description
NEXTCLOUD_URL	Yes	—	Nextcloud base URL (e.g., https://cloud.example.com)
NEXTCLOUD_USER	Yes	—	Bot username
NEXTCLOUD_PASSWORD	Yes	—	Bot password (or app password)
NEXTCLOUD_WEB_PASSWORD	No	same as PASSWORD	Separate browser login password (Yunohost SSO)
GEMINI_API_KEY	Yes	—	Google Gemini API key
AUTH_METHOD	No	nextcloud	nextcloud (standard login) or yunohost (SSO)
GEMINI_MODEL	No	gemini-2.5-flash-lite	Gemini model for transcription
NOTES_STORAGE	No	nextcloud	nextcloud (WebDAV) or local (filesystem)
NOTES_FOLDER	No	/Talk/Notes	Nextcloud folder for notes
LOCAL_NOTES_DIR	No	—	Local directory for notes (when storage=local)
FILENAME_LAST_USER	No	—	User ID always placed last in filenames
POLL_INTERVAL_SECONDS	No	10	Seconds between poll cycles
AUDIO_DIR	No	/tmp/notetaker-audio	Temp directory for audio files
SMTP_HOST	No	localhost	SMTP server
SMTP_PORT	No	25	SMTP port
SMTP_FROM	No	—	Sender email address
SMTP_USER	No	—	SMTP username
SMTP_PASSWORD	No	—	SMTP password
EMAIL_OVERRIDES	No	—	Comma-separated user=email overrides
MAIL_DOMAIN	No	—	Fallback domain for email addresses

Notes Output

Each call produces a markdown file with:

• Meeting metadata (date, duration, participants)
• Topics discussed with detailed summaries
• Action items table with owners and deadlines
• Decisions made
• Open questions and follow-ups
• Follow-up email draft

Architecture Notes

• Audio capture: In-browser WebRTC interception via Playwright. No PulseAudio needed.
• Transcription: Chunked pipeline — long recordings (>10 min) are split into 15-minute MP3 segments, each transcribed separately, then synthesized into final notes.
• Muting: The bot uses synthetic silent audio/black video streams so other participants don't hear Chrome's fake device beep.
• Gemini cost: ~$0.02 per 30-minute call on Flash Lite. Free tier (1000 req/day) is sufficient for most teams.

Troubleshooting

Bot doesn't join calls unless it is added as a participant: Make sure the bot user is added as a participant in the Talk room.

"Could not connect" during setup: Verify the Nextcloud URL is correct and starts with https://.

"401 Unauthorized": Wrong bot username or password.

"404 Not Found" on Talk API: The Nextcloud Talk app may not be installed.

No email notifications: The bot needs subadmin status for user groups to look up email addresses. See setup instructions above.

Transcription quality: Gemini may mishear proper nouns. This is inherent to speech-to-text. For important terms, check the notes after generation.

Development

python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"
playwright install chromium
pytest

License

AGPL-3.0 — same as Nextcloud.