CushLabs Messenger
The 24/7 bilingual AI assistant for Facebook Messenger — built from your business's own content, so no lead goes cold.
Video Walkthrough

Screenshots
1 / 10









The Challenge
Businesses that market on Facebook get Messenger questions about pricing, hours, and services at all hours, in whatever language the customer speaks. Nobody watches the inbox around the clock, customers expect fast answers, and a generic auto-reply — or a chatbot that invents details — costs real business. The usual agency fix is a separate bot build per client: separate deploy, separate Meta app, separate review cycle, separate thing to maintain. That does not scale past a couple of clients.
The Solution
One Cloudflare Worker serves every connected Page. Each inbound webhook is routed by page_id to that tenant's own system prompt, per-Page access token, structured business facts, and tenant-scoped vector index — so onboarding a client is a config and content operation, not a deploy. A two-layer router sends pleasantries to Claude Haiku and real questions to Claude Sonnet with RAG context. Hard facts (hours, pricing, address, booking URL) live as structured records injected verbatim and marked authoritative, so a price change is a one-field edit that is live on the next message with no re-embed. When a customer wants a person, the bot goes quiet and alerts the operator, who replies from the Page Inbox.
Key Features
- Answers customer questions in under 5 seconds, 24/7 — nights, weekends, holidays
- English and Spanish automatically, following whatever the customer writes, with no manual switching
- Answers come from the client's own approved content — it never invents a price or makes up a promise
- Hours and pricing come back exact every time, because they are stored as data, not as prose the AI has to interpret
- Hands off to a person with full context when a customer asks, alerts the owner, then resumes on its own
- One Worker, many client Pages — a new client onboards as configuration, with no code change and no Meta re-review
Results
- 2 live client Pages running on a single Worker deployment
- 485 automated tests across 29 files, covering the decision layer and not just API plumbing
- 53 scripted QA scenarios run against production before a client goes live, cross-tenant isolation included
- ~50 seconds from merge to live, typecheck and tests gated
Related Projects
Converso AI — Bilingual AI Front Desk & Sales Assistant
Bilingual AI chatbot platform for service businesses — handles front desk and sales conversations 24/7
NY AI Chatbot
AI chatbot for a coaching business — turns after-hours website visitors into booked calls
MarketSignal — Google Review Management & Local Competitive Intelligence
Google review management and competitor tracking — AI-drafted replies plus one weekly action item, delivered to WhatsApp
Have a similar challenge?
Let's talk about how a solution like this could work for your team. Free call, no pitch.
Book a Free Call
