millan-amara/agent

whatsapp native autopilot crm # Azayon WhatsApp-first AI lead system for SMEs. Businesses connect their WhatsApp number; the AI auto-replies in their brand voice, qualifies leads, updates the CRM, schedules follow-ups, and hands off to humans. Full product/architecture spec: PLAN.md. This repo currently implements **Slice 1**: the message loop — webhook → queue (per-contact ordering + debounce) → Claude agent with CRM tools → window-aware sender — plus a CLI simulator. ## Layout ``` apps/api Fastify backend (TypeScript, ESM) prisma/ Schema — SQLite in dev, Postgres in production src/ agent/ System-prompt builder, CRM tools, the agent loop whatsapp/ Cloud API webhook (verify + HMAC) and senders queue/ Debounced per-contact FIFO (in-memory dev driver) followups.ts Due-follow-up worker (window-aware) simulator.ts CLI chat against the real agent loop ``` ## Setup ```sh npm install cp apps/api/.env.example apps/api/.env # then set ANTHROPIC_API_KEY npm run db:push -w apps/api # creates apps/api/prisma/dev.db ``` ## Run **Simulator (no Meta credentials needed):** ```sh npm run simulator ``` Chat as a customer of the seeded dev tenant (a Nairobi physio clinic). `/lead` shows the CRM record the AI builds as you talk; `/reset` starts a new customer. **Server (webhook mode):** ```sh npm run dev ``` Requires `WA_ACCESS_TOKEN`, `WA_PHONE_NUMBER_ID`, `WA_APP_SECRET` in `.env`, and the webhook URL (`/webhooks/whatsapp`) registered in the Meta app dashboard (use a tunnel like cloudflared/ngrok locally). ## Dev vs production | Concern | Dev (this machine) | Production | |---|---|---| | DB | SQLite (`file:./dev.db`) | Postgres + tenant RLS | | Queue | In-process debounced FIFO | BullMQ + Redis (same interface) | | Outbound | Console / Cloud API | Cloud API + template messages | | Tenancy | Hardcoded dev tenant | Embedded Signup onboarding |