Parley

AI meeting notes that cite their sources. An open-core clone of Fathom, built in 24 hours on Next.js, Vercel and Neon Postgres.

• Live: https://parley-smoky.vercel.app
• Handover guide (how to use it, and how it compares with Fathom): HANDOVER.md · illustrated PDF

Parley records Zoom, Google Meet and Teams calls from your browser: your mic plus the meeting tab's audio, with no bot joining the call. It transcribes live with Deepgram and writes notes while you talk. When the call ends you get a summary, action items with owners, clips and a Slack briefing. Every note links to the exact line of the transcript it came from.

Features

Area	What it does
Recording	Mic + meeting-tab audio in the browser, pause/leave/discard, consent notice, audio archive with seekable playback
Transcription	Deepgram nova-3 streaming with speaker diarization; typed-line fallback without a key
Live notes	Decisions, action items and questions during the call: AI on closed 10-minute windows and on the open end (refreshed ~30s), rule-based for the newest lines
After the call	Summary in your template, action items with owners and due dates, highlights, shareable clips; every item cites its transcript line
Ask Parley	Chat across all meetings and your calendar; every sentence cited; uncited sentences are dropped
Search	Full-text across calls with synonyms (SSO = single sign-on = SAML), ⌘K command bar
Scratchpad	Private per-meeting notes, stamped to the minute, never sent to AI
Calendar	Google/Outlook/iCloud via a secret iCal address (no OAuth); join-and-record from upcoming events
Integrations	Slack briefing, Notion page per meeting, HubSpot note on matched contacts; each user's own tokens
Models (BYOK)	Claude, ChatGPT or Hugging Face, from the user's key or the server's, tried in that order with fallthrough
Developer	REST API (/api/v1) and a read-only MCP server (/api/mcp) with personal access tokens
Accounts	Try now (no signup, deleted after 24 h), Sign in with Google, shared reviewer demo
Data controls	Export everything (JSON), retention (30/90/365 days), delete account, security activity log
Onboarding	Guided tour with a gradient cursor on the landing page and dashboard

Quick start

npm install
cp .env.example .env.local     # set DATABASE_URL at minimum, or use pglite:./.pglite
npm run db:migrate
npm run db:seed                # demo workspace: Maya Chen, 20 meetings
npm run dev                    # http://localhost:3000

No cloud database needed locally: DATABASE_URL=pglite:./.pglite (also set DATABASE_URL_UNPOOLED to the same value for db:migrate) runs an embedded Postgres.

Environment variables

Variable	Required	Purpose
DATABASE_URL	yes	Neon pooled connection string (or pglite:./.pglite)
DATABASE_URL_UNPOOLED	yes	Neon direct connection, used by migrations
PARLEY_SECRET_KEY	yes	32+ random chars (openssl rand -base64 32). Signs sessions, encrypts stored keys. Never rotate after launch.
CRON_SECRET	production	Protects /api/cron/sweep (openssl rand -hex 32)
APP_URL	production	Public base URL, no trailing slash. Used for links in Slack/Notion/HubSpot/API
GOOGLE_CLIENT_ID, GOOGLE_CLIENT_SECRET	for accounts	Google OAuth web client; redirect URI <APP_URL>/api/auth/google/callback
DEEPGRAM_API_KEY	recommended	Live transcription (key role Member or higher)
ANTHROPIC_API_KEY, ANTHROPIC_MODEL	optional	Default model for everyone (claude-opus-5)
OPENAI_API_KEY, OPENAI_MODEL	optional	Fallback model (gpt-5-mini)
HF_TOKEN, HF_MODELS	optional	Free Hugging Face models, tried in order
SLACK_WEBHOOK_URL	optional	Server-wide briefing channel, demo workspace only

Users' own keys (Settings) always win over the server's. Notion and HubSpot have no server-wide key: they're per user.

Commands

Command	What it does
npm run dev / build / start	Next.js dev server / production build / serve
npm run typecheck	tsc --noEmit
npm run db:migrate	Apply drizzle/*.sql to DATABASE_URL_UNPOOLED (or DATABASE_URL)
npm run db:seed	Load the demo fixtures (idempotent)
npm run db:verify	Schema, AI pipeline and open-core suites against an in-memory Postgres. No network or keys needed
npm run verify:routes	HTTP-level suite against a running server (BASE_URL, default http://localhost:3100). Run the server without HF_TOKEN, on PGlite

API

Create a personal access token in Settings → Access tokens with the read scope and send Authorization: Bearer parley_pat_…. Limit: 300 requests per 5 minutes.

Method	Path	Returns
GET	/api/v1/meetings?limit=	Recent meetings with overview, participants, action-item counts
GET	/api/v1/meetings/:id?transcript=1	Summary sections, action items, highlights, clips, optional transcript
GET	/api/v1/action-items?status=open|done	Action items with owner, due date, source link
GET	/api/v1/search?q=	Transcript moments and calendar events (synonym-aware)
POST	/api/v1/ask {"question": "…"}	Cited answer across meetings and calendar

MCP server

Stateless Streamable HTTP at /api/mcp (protocol 2025-11-25, also 2025-06-18, 2025-03-26, 2024-11-05). Read-only tools: search_meetings, ask_parley, list_meetings, get_meeting, list_action_items.

claude mcp add --transport http parley https://parley-smoky.vercel.app/api/mcp --header "Authorization: Bearer parley_pat_…"

Architecture

Browser ── mic + tab audio ──▶ Deepgram (WebSocket, 60s token from /api/deepgram/token)
   │                                   │ words + speakers
   ▼                                   ▼
/api/ingest (append lines) ──▶ Postgres (Neon) ◀── /api/streams/meetings (SSE to watchers)
   │ after(): job queue                 ▲
   ▼                                    │
AI pipeline: 10-min windows ▸ merge ▸ grounded summary ▸ Slack / Notion / HubSpot
   (Claude → ChatGPT → Hugging Face, user key first)
Vercel Cron (daily) ▸ /api/cron/sweep: stale calls, stuck jobs, iCal sync, retention, guest expiry

• Stack: Next.js 16 (App Router, after()), React 19, TypeScript, Tailwind v4, Drizzle ORM, Neon serverless (PGlite locally), Vercel.
• Grounding: the model only sees numbered transcript lines and must cite them. Citations that point nowhere, and quotes that don't match the line, are dropped or flagged.
• Security: HMAC-signed session cookies; AES-256-GCM for stored keys and feed URLs; SHA-256-hashed access tokens with scopes (ingest, calendar, read); Postgres sliding-window rate limits; owner checks on every write.
• Storage: everything lives in Postgres, including audio chunks (~15-20 MB per hour). Move audio to Vercel Blob or R2 before it outgrows Neon's free 0.5 GB.

Project layout

app/                 routes: landing, (app) dashboard pages, api/*, server actions
src/ai/              LLM clients + pool, pipeline, prompts, grounding, Ask Parley
src/calendar/        iCal fetch/parse/sync
src/capture/         dual-stream (mic + tab) capture, recorder
src/notify/          Slack, Notion, HubSpot delivery
src/db/              schema, seed, client (Neon or PGlite)
src/components/      UI (shadcn), live call, tour, settings forms
drizzle/             SQL migrations 0000-0011
scripts/             migrate + verification suites
seed/                demo fixtures (20 meetings)
.agent-logs/         build session logs (kept in the repo on purpose)

Deploy

npx vercel link
npx vercel env add DATABASE_URL production     # repeat for each variable above
npx vercel deploy --prod

vercel.json pins functions to cle1 (next to Neon us-east-2) and schedules the daily sweep.

Limits

No meeting bot, no video recording, no Google Calendar OAuth, no SSO/SOC 2/HIPAA. HANDOVER.md explains why each was left out.