English · Deutsch | Website: wortlaut.cloud · Docker Quickstart: docker-compose.yml · Helm Chart: charts/wortlaut · Configuration Guide: docs/configuration.md · Changelog: CHANGELOG.md · Security: SECURITY.md
Speaker-accurate meeting transcripts and protocols for Google Meet, Microsoft Teams, Cisco Webex, and Zoom — 100% self-hosted on your own infrastructure.
wortlaut joins online calls as a transparent browser participant, captures each speaker’s isolated WebRTC audio stream before conference mixing, and transcribes everything locally on your own GPU or CPU — with automatic calendar invite writeback, live in-call chat opt-out, local LLM summaries, and .md / .txt / .vtt / .json exports.
Commercial meeting bots (Fireflies, Otter, Fathom, Read.ai) route confidential conversations through third-party clouds, build biometric voiceprints, and force attendees into an all-or-nothing choice: accept full cloud recording or leave the call.
wortlaut is built for privacy-conscious engineering teams, consultancies, enterprises, and public-sector organizations:
| What teams actually need | How wortlaut solves it |
|---|---|
| Zero third-party cloud leakage | Speech-to-text (Qwen3-ASR 1.7B — 30 languages with automatic language detection, 3.88% German WER on the Open ASR Leaderboard) and optional summaries (Ollama) run 100% inside your own cluster or server. |
| True speaker attribution — no guessing | Instead of recording mixed audio and guessing speakers with brittle voice diarization, wortlaut taps isolated per-participant WebRTC receiver tracks (RTCRtpReceiver) before mixing. |
| Individual opt-out right in the call | Any participant who prefers not to be quoted simply types opt out or no (nein) in the meeting chat. The call continues seamlessly for everyone else while their track and prior statements are purged immediately. |
| Protocols right where people look | As soon as transcription finishes, wortlaut writes a read-only transcript link directly back into the Google Calendar or Microsoft 365 / Outlook event, emails the protocol, and offers .md, .txt, .vtt, and .json downloads. |
| Click-to-listen verification | Every statement has a clickable timestamp. During the 24-hour raw-audio retention window, operators can play back the exact isolated audio snippet in the browser before raw audio is automatically purged. |
| Screen-share slide keyframes (never camera video) | Captures settled static PNG snapshots (<= 1280 × 720) strictly from verified screen-share streams (WORTLAUT_ENABLE_KEYFRAMES=true) — participant camera feeds are rejected fail-closed. Slides appear inline on the transcript timeline and via MCP without recording creepy video. |
- Automatic Calendar Sync (Google Calendar & Microsoft 365):
Connect your calendar once via OAuth2 in the Web-UI (
/ui#view=calendar). Self-organized meetings with confirmed guests auto-join automatically; external invitations can be toggled on or off with 1 click. - Instant Ad-Hoc Link:
Paste any Google Meet, Microsoft Teams, Cisco Webex, or Zoom URL on the home screen —
wortlautjoins within ~15 seconds. - Email /
.icsInvitation: Invite the bot email address to any calendar invite like a regular attendee, or upload an.ics/.emlfile directly.
- Upon joining,
wortlautdisplays the inviting person’s name — e.g.wortlaut (Elena Rostova)— renders live on-screen chat instructions directly on its1280×720WebRTC video tile (video-tile.ts), and posts a concise notice in the meeting chat with a link to your instance’s Privacy & Transparency Notice (/privacy) (explicitly noting whenWORTLAUT_ENABLE_KEYFRAMES=truethat participant camera video is never recorded and only settled screen-share snapshots are captured; see./docs/configuration.md). - Live Chat Opt-Out: If any participant writes
no,opt out,optout, orstopin the meeting chat during the call, their audio stream and screen-share sampler are halted on the spot and any already-transcribed segments and uploaded screen-share keyframes from that participant are purged immediately from PostgreSQL and S3 (meet-page.ts, processing.py). - Off-the-Record Pause (
#pause/#resume): Typing#pause(or/pause,!pause,#off-the-record,#pausieren) in the meeting chat immediately zeroes audio capture in worker RAM and suspends screen-share keyframe sampling while switchingwortlaut’s video tile to a high-visibility amber OFF THE RECORD indicator. Typing#resume(or#weiter,/resume) resumes recording without leaving the room (capture.ts, meet-page.ts). - Live Audio Activity Indicator: A real-time VU waveform meter on
wortlaut’s16:9video tile and in the Web-UI shows immediately when audio is actively entering the recording pipeline.
- Automatic Calendar Writeback: Once completed,
wortlautappends the scoped read-only protocol link (wortlaut-read-…) directly into the Google Calendar or Microsoft 365 invite description (calendar_oauth.py). - Bilingual Web-UI (
EN/DE) & Protocol Editor: Switch between English and German at any time; search across all transcripts, rename guest speakers (speaker_overrides), and apply 1-click domain vocabulary replacements across the entire protocol. - Exports for Every Workflow: Copy GitHub-Flavored Markdown (
.md) straight into Notion, Confluence, GitHub, GitLab, or Jira, download.txt,.vtt, or.json, or share a zero-account read-only link (wortlaut-read-…) (exports.py).
All four platforms work side-by-side in the same wortlaut deployment:
| Platform | Join Mode & Capabilities |
|---|---|
| Google Meet | Joins as guest or with a saved Google session (meet.ts, meet-page.ts, capture.ts); isolated WebRTC receiver tracks, participant binding, 1280×720 (16:9) live transcript/activity video tile (video-tile.ts), settled screen-share keyframes (a=content:slides + screencast DOM, camera tracks excluded), and live chat controls (no/opt out, #pause/#resume). |
| Microsoft Teams | Joins via Teams web client in English or German (teams-browser.ts); isolated WebRTC receiver capture, CSI speaker binding, 1280×720 (16:9) live video tile, applicationsharing-video screen-share keyframes, and live chat controls (no/opt out, #pause/#resume). |
| Cisco Webex | Joins as browser guest (webex-browser.ts) or via Service App SDK; multistream AUDIO-MAIN capture, activeSpeakerNotification binding, 1280×720 (16:9) live video tile + VIDEO/SLIDES screen-share keyframes in browser mode ("unsupported" in Guest-Issuer SDK mode), and live chat controls (no/opt out, #pause/#resume). |
| Zoom | Joins via Zoom Web Client as guest or authenticated session (zoom-browser.ts, zoom-url.ts); waiting-room handling, Opus capture, 1280×720 (16:9) live video tile, .sharing-layout-container / #sharee-container-canvas screen-share keyframes bound to a unique roster presenter, and live chat controls (no/opt out, #pause/#resume). |
(Whenever a conference provider does not expose a verified participant binding for a stream, wortlaut surfaces attribution_unknown: true transparently instead of guessing speaker names.)
wortlaut supports four runtime backends (runtime_backends.py) and runs with either an NVIDIA GPU (recommended for real-time throughput) or CPU fallback:
Starts wortlaut-meeting (non-root, isolated via docker-socket-proxy), wortlaut-asr, PostgreSQL 17, and RustFS S3 (docker-compose.yml):
cp .env.example .env
# Set WORTLAUT_API_TOKEN and passwords in .env, then launch:
docker compose up -d
# Recommended when an NVIDIA GPU is available (real-time Qwen3-ASR):
docker compose -f docker-compose.yml -f docker-compose.gpu.yml up -d
# Open Web-UI: http://127.0.0.1:34601/uiZero-dependency evaluation install for K3s, kind, or homelabs (automatically provisions embedded PostgreSQL 17, RustFS S3, and persistent random Secrets):
helm upgrade --install wortlaut oci://ghcr.io/mayflower/charts/wortlaut \
--namespace wortlaut --create-namespace \
--set host=wortlaut.local \
--set bundled.enabled=true \
--set asr.gpuCount=0 --set asr.allowCpu=trueProduction cluster install with external PostgreSQL, S3, GPU node scheduling, and OIDC SSO (full guide in charts/wortlaut/README.md and docs/configuration.md):
helm upgrade --install wortlaut oci://ghcr.io/mayflower/charts/wortlaut \
--namespace wortlaut --create-namespace \
--set host=wortlaut.example.org \
--set oidc.issuer=https://auth.example.org./scripts/run-local.shThe wortlaut-asr service (services/asr/app.py) loads Qwen/Qwen3-ASR-1.7B-hf at a pinned revision through Hugging Face Transformers (WORTLAUT_ASR_MODEL_REPO and WORTLAUT_ASR_MODEL_REVISION override both). All 30 Qwen3-ASR languages are enabled out of the box (WORTLAUT_ASR_LANGUAGES=*: Chinese, English, Cantonese, Arabic, German, French, Spanish, Portuguese, Indonesian, Italian, Korean, Russian, Thai, Vietnamese, Japanese, Turkish, Hindi, Malay, Dutch, Swedish, Danish, Finnish, Polish, Czech, Filipino, Persian, Greek, Hungarian, Macedonian, Romanian) and can be selected or switched per recording in the Web-UI, via POST/PATCH /v1/recordings, or via MCP. Operators can optionally set WORTLAUT_ASR_LANGUAGES (e.g. German,English) to define a narrower default for new recordings.
The model returns no word timestamps, so each ASR job receives only its own disjoint audio span and transcript segments carry segment-level times. Word confidence is the mean probability of the generated tokens. Silence and noise produce an empty transcript; output that reaches the generation budget without an end marker is discarded as a decoder loop. Set WORTLAUT_ASR_GLOSSARY="TermA, TermB" (comma-separated or JSON array) on the meeting service to bias Qwen3-ASR system context prompts alongside meeting titles and confirmed participant names. Active meetings that run past their scheduled ends_at automatically extend in 15-minute increments up to WORTLAUT_MAX_OVERTIME_SECONDS (default 3600 = 60 minutes).
Set WORTLAUT_DEFAULT_LOCALE=en (already enabled in .env.example and docker-compose.yml) to default the Web-UI, /privacy page, and in-call chat announcements to English.
- Short Raw-Audio Window (24 Hours): Isolated speaker audio chunks are used strictly for asynchronous ASR and short snippet verification, then automatically deleted from S3 24 hours after the meeting ends.
- Automatic Transcript & Slide Keyframe Expiration (30 Days): Transcripts, settled screen-share keyframes (
keyframes/{recording_id}/), and metadata share theWORTLAUT_TRANSCRIPT_RETENTION_DAYSlifecycle (30 days by default in settings.py, counted towardWORTLAUT_MAX_STORED_AUDIO_BYTES) and can be deleted immediately as a whole (DELETE /v1/recordings/{id}) or slide-by-slide (DELETE /v1/recordings/{id}/keyframes/{keyframe_id}). - Multi-Tenant Isolation & Scoped Share Tokens: Users signed in via OIDC (oidc.py — Dex, Keycloak, Authentik, Zitadel, Entra ID, Google) only see their own recordings and calendar events they were invited to. Shared
wortlaut-read-<recording_id>.<hmac>links grant read-only access to a single recording.
wortlaut includes a built-in MCP Streamable HTTP server (POST /mcp, mcp_server.py) protected by OAuth 2.1 (mcp auth) with RFC 9728 Protected Resource Metadata, RFC 8414 Authorization Server Metadata, RFC 7591 Dynamic Client Registration, PKCE S256, and explicit user consent (mcp_oauth.py):
claude mcp add --transport http wortlaut https://wortlaut.example.org/mcp- Zero static API keys: MCP clients (
Claude Code,Cursor,VS Code,Antigravity) automatically discover OAuth endpoints, sign users in via your existing OIDC provider (or local instance token), and requestmcp:readand optionalmcp:writescopes. - Strict per-user isolation & revocation: Every tool call (
list_recordings,get_recording,get_transcript,export_protocol,list_calendar_events,schedule_recording,cancel_recording,update_protocol,generate_summary,toggle_calendar_event_recording,publish_calendar_protocol_link) andwortlaut://resource (including screen-share slide snapshots via thewortlaut://recordings/{recording_id}/keyframes/{keyframe_id}.pngresource template) is scoped to the authenticated user. Active client grants can be reviewed and revoked at/ui#view=mcp(GET/DELETE /v1/mcp/grants).
Point wortlaut at a self-hosted Ollama instance to generate structured summaries, topic breakdowns, and action items with zero external API calls:
WORTLAUT_ENABLE_LLM_SUMMARY=true
WORTLAUT_OLLAMA_URL=http://ollama.ollama.svc.cluster.local:11434
WORTLAUT_OLLAMA_MODEL=qwen3.8:latestReceive .ics calendar invitations (with DKIM verification) and automatically email finished transcripts (mail.py, calendar.py):
WORTLAUT_ENABLE_IMAP=true
WORTLAUT_IMAP_HOST=imap.example.org
WORTLAUT_IMAP_USERNAME=bot@example.org
WORTLAUT_IMAP_PASSWORD=REPLACE_PRIVATELY
WORTLAUT_MAIL_SENDER_AUTH=dkim
WORTLAUT_MAIL_ALLOWED_SENDER_DOMAINS=["example.org"]
WORTLAUT_MAIL_REPLY_TO_SENDER=true
WORTLAUT_ENABLE_SMTP=true
WORTLAUT_SMTP_HOST=smtp.example.org
WORTLAUT_SMTP_USERNAME=bot@example.org
WORTLAUT_SMTP_PASSWORD=REPLACE_PRIVATELY
WORTLAUT_SMTP_FROM=bot@example.orgwortlaut consists of three focused components:
- Meeting Service (services/meeting, app.py, settings.py) — Multi-arch (
linux/amd64,linux/arm64) Python/FastAPI service hosting the REST API, embedded Web-UI (ui.py), OIDC PKCE auth (oidc.py), Google/M365 Calendar OAuth2 sync (calendar_oauth.py), scheduler, Kubernetes/Docker worker controller (controller.py), audio ingest, exports (exports.py), and IMAP/SMTP mail (mail.py). - Browser Worker (worker) — Ephemeral TypeScript + Playwright + sandboxed Chromium job/container spawned per meeting to capture isolated pre-mix WebRTC receiver tracks.
- ASR Service (services/asr, services/asr/app.py) — Shared local Qwen3-ASR speech recognition service (
WORTLAUT_ASR_MODEL_REPO,WORTLAUT_ASR_MODEL_REVISION,WORTLAUT_ASR_ALLOW_CPU).
| Endpoint | Purpose |
|---|---|
POST /v1/recordings |
Schedule or immediately start a meeting recording (with Idempotency-Key) |
GET /v1/recordings |
List recordings & search full text or speaker names (?q=) |
GET /v1/recordings/{id} |
Recording status, track diagnostics (live_status, live_status.paused), and opted_out_speakers |
POST /v1/recordings/{id}/cancel / /rejoin |
Stop an active/scheduled recording or dispatch a fresh worker to rejoin an ongoing meeting |
PATCH /v1/recordings/{id} |
Update title, notes, speaker_overrides, term_replacements, action items, or opted_out_speakers |
DELETE /v1/recordings/{id} |
Immediately purge recording, transcript, S3 audio chunks, and slide keyframes |
GET /v1/recordings/{id}/transcript |
Retrieve speaker-attributed segments, word confidences, keyframes, and pause_windows |
GET /v1/recordings/{id}/exports/{format} |
Download protocol as md, txt, vtt, or json |
GET /v1/recordings/{id}/audio-snippet |
Stream isolated WAV snippet for statement verification (24h window) |
GET /v1/recordings/{id}/keyframes/{kf_id}.png (DELETE …/{kf_id}) |
Fetch (Authorization: Bearer) or permanently delete a settled screen-share slide PNG |
POST /v1/recordings/{id}/summary |
Generate local LLM executive summary & action items via Ollama |
POST /v1/recordings/{id}/calendar-link |
Write protocol link into the linked Google Calendar / M365 event |
POST /v1/recordings/{id}/reader-token |
Issue scoped read-only wortlaut-read-… share token |
POST /v1/calendar/messages |
Import .ics or RFC822 .eml calendar invitations |
GET /v1/calendar |
List connected OAuth2 calendars, synced events, and .ics invitations |
POST /v1/calendar/oauth/exchange |
Connect Google Calendar (calendar.events) or M365 (Calendars.ReadWrite) via PKCE |
POST /v1/calendar/oauth/events/{id}/toggle |
Enable or disable auto-recording for a synced calendar event |
POST /mcp (GET/DELETE /v1/mcp/grants) |
Authenticated MCP Streamable HTTP server (OAuth 2.1 mcp:read / mcp:write) & grant revocation |
GET /privacy / GET /metrics |
Public privacy notice (?lang=en / ?lang=de) and Prometheus metrics |
IDEMPOTENCY_KEY=$(uuidgen)
jq -n --arg url "$MEETING_URL" --arg start "$STARTS_AT" --arg end "$ENDS_AT" \
'{provider:"google_meet", meeting_url:$url, starts_at:$start, ends_at:$end}' |
curl --fail-with-body http://127.0.0.1:34601/v1/recordings \
-H "Authorization: Bearer $WORTLAUT_API_TOKEN" \
-H 'Content-Type: application/json' \
-H "Idempotency-Key: $IDEMPOTENCY_KEY" --data-binary @-
# Download finished Markdown protocol:
curl --fail-with-body "http://127.0.0.1:34601/v1/recordings/$RECORDING_ID/exports/md" \
-H "Authorization: Bearer $WORTLAUT_API_TOKEN" -o protocol.mdExternal contributors only need Python 3.12 (uv), Node.js 22+, and Docker — no Kubernetes cluster required (see CONTRIBUTING.md):
make setup-local # Install Python & Node dependencies, build bundles, and install git QA hooks
make qa # Run fast secret, lint, typecheck, and Helm/workflow QA checks
make test-local # Run linters + full PostgreSQL, S3, and worker integration tests in disposable containersAll configuration variables and defaults are documented in settings.py and docs/configuration.md. Please report security vulnerabilities privately via SECURITY.md.

