Switchboard

Self-hosted, multi-tenant omnichannel customer-engagement platform.

Unify customer conversations from across many messaging accounts into one real-time team console — with role-based collaboration, full audit trails, embedded real voice calls, and AI assist.

Next.js 16 · React 19 · TypeScript · PostgreSQL · Prisma · WebRTC

---

Switchboard turns scattered messaging accounts into a single collaborative inbox for a support team. Agents see every customer conversation in one place, claim and route them, reply in real time, place voice calls without leaving the browser, and leave a complete audit trail behind every action. Each Workspace is an isolated tenant with its own accounts, members, roles, and settings.

Telegram is the first fully-implemented channel. The data model (Account → Channel → Conversation → Message) and the bridge worker are channel-agnostic by design, so additional providers slot in behind the same pipeline and UI.

Why it's worth a look (engineering highlights)

• 🎙️ Embedded real voice calls — a WebRTC ⇄ Telegram MTProto bridge lets agents place and answer real Telegram voice calls from the browser. A Node gateway handles signalling; a Python media helper bridges py-tgcalls/ntgcalls (the real, encrypted TG call) to aiortc with full-duplex PCM. No audio is persisted.
• ⚡ Real-time by default — server-sent events over an in-memory event bus with automatic reconnect and lastEventId replay (5-minute ring buffer), so flaky mobile/corporate networks recover transparently.
• 🏢 Multi-tenant from the ground up — strict per-workspace isolation and a Discord-style RBAC layer (custom roles, granular permission keys, permissions unioned across roles).
• 🔐 Security-first — AES-256-GCM encryption of stored provider sessions, bcrypt auth with timing-attack mitigation, per-account login throttling, a runtime secret-guard that refuses to boot on placeholder secrets, strict security headers + HSTS, and access-controlled media.
• 🤖 AI assist — pluggable LLM provider for intent/sentiment/urgency analysis of inbound messages.
• 📱 Mobile-first PWA — installable, offline-aware, safe-area-padded, ≥44px tap targets; built for agents on phones.

Features

Area	What it does
Unified inbox	Multi-account send/receive, pinning, edit history, emoji reactions, history backfill
Conversation routing	Claim/assign conversations, ownership, snooze/close lifecycle
Voice calls	Place/answer real Telegram 1:1 voice calls in-browser
AI analysis	Intent, sentiment, and urgency hints on inbound messages
Quick replies	/shortcut-triggered reply templates
Roles & permissions	Custom roles with granular, unioned permission keys
Scheduling	Time-based rules for routing/automation
Audit log	Every key operation recorded, retained, and queryable
Multi-tenant	One deployment, many isolated teams

Architecture

 Messaging accounts ┐
   (Telegram first) │   ┌─────────────────────────┐  HTTP    ┌──────────────────────┐
   many accounts ───┼──▶│  Bridge worker  :3001    │─────────▶│  Next.js app  :1688  │
                    │   │  GramJS client pool      │          │  API · SSE · UI       │
                    │   │  message pipeline        │◀─────────│  realtime push        │
                    ┘   └─────────────────────────┘  notify  └──────────┬───────────┘
                                                                         │
            ┌──────────────────────────────┐                 ┌──────────▼─────────┐
            │ Voice gateway :3002 + media   │                 │   PostgreSQL 16    │
            │ helper :3003 (WebRTC ⇄ MTProto)│                 └────────────────────┘
            └──────────────────────────────┘

Two long-running processes are required: the app (npm run dev, port 1688) and the bridge worker (npm run bridge, port 3001). Voice calls add the gateway + Python media helper.

Tech stack

Layer	Technology
Framework	Next.js 16 (App Router, Turbopack)
Language	TypeScript 5.9 (strict)
UI	React 19 · Tailwind CSS 4 · Lucide
Database	PostgreSQL 16
ORM	Prisma 6.19
Auth	Custom session (bcryptjs + HTTP-only cookie)
Messaging	GramJS (Telegram MTProto)
Voice	py-tgcalls / ntgcalls ⇄ aiortc
Encryption	AES-256-GCM (Node crypto)

Quick start

cp .env.example .env          # fill in values (see below)
docker compose up -d          # PostgreSQL
npm install
npm run db:migrate            # apply schema
npm run db:generate           # Prisma client
SEED_ADMIN_PASSWORD=... SEED_CS_PASSWORD=... npm run db:seed
npm run dev                   # app  → http://localhost:1688
npm run bridge                # bridge worker (separate terminal)

Required environment (generate secrets with openssl rand -hex 32)

Variable	Purpose
DATABASE_URL	PostgreSQL connection string
SESSION_SECRET	session integrity (required in production)
TELEGRAM_SESSION_KEY	AES-256-GCM key for stored provider sessions
INTERNAL_SECRET	shared secret for app ↔ bridge
SEED_ADMIN_PASSWORD / SEED_CS_PASSWORD	required when seeding production
TRUSTED_PROXY_HOPS	trusted reverse-proxy depth (rate-limit/audit IP)

Per-account Telegram apiId/apiHash are entered through the in-app account UI (from https://my.telegram.org), never via environment variables.

See SECURITY.md before deploying.

Notes

• This project automates Telegram user accounts (GramJS) and bridges real calls; review Telegram's Terms of Service and your local data-protection law before deploying it against real accounts.
• The voice dependencies (py-tgcalls, ntgcalls, telethon, aiortc) carry their own licenses — review them before redistribution.

License

MIT