Schemes Broadcast MCP

A B2B tool for FMCG distributors. The distributor's own Claude or ChatGPT drafts combo offers and discounts through MCP tools, the distributor approves them on a web dashboard, and the app generates a public shareable link that buyers (kirana retailers) open with no login.

There is no AI/LLM API key anywhere in this codebase. All generation happens in the distributor's own Claude/ChatGPT session. This app is a plain CRUD backend + an MCP tool surface + a dashboard + a public offer page.

Live demo: https://schemes-broadcast-mcp.vercel.app (sign up with any email to try it)

Why this exists

Distributors announce schemes ("10% off Tata Salt this week") by phone calls and unstructured WhatsApp text. There is no record, no targeting, and no safe approval step. Claude alone cannot persist an offer, show a tap-to-approve card, or give a retailer without Claude a link to open. This app does exactly those parts. See the Approval flow section below.

Related MCP server: E-commerce MCP Server

How it works

Distributor's Claude/ChatGPT --(MCP tools, personal token)--> /api/mcp --> Supabase
                                                                     |
Distributor dashboard (login) <--- pending offers ---------------------+
        | Approve (edit if needed)
        v
public link  /o/<slug>  --> buyer opens it, taps "I'm interested" --> counted on dashboard

MCP tools

Approval flow

create_offer only ever inserts status = 'pending' and does not create a share link. The distributor opens the offer card in the dashboard, edits any field, and clicks Approve. Only then is a random link slug generated and the offer becomes visible at /o/<slug>. Rejected or pending offers show "This offer is no longer available."

Security model

• Each distributor has a personal MCP token. The server resolves it to exactly one distributor_id and scopes every read/write to it. create_offer also checks that the product ids and group id belong to that distributor.

• The dashboard uses a Supabase session with Row Level Security policies (see supabase/schema.sql) plus explicit distributor_id filters.

• The public page and the interest endpoint use the service-role key server-side only, and match only status = 'approved' offers. The service-role key is never sent to the browser.

• The token can be regenerated from the Connect page (this invalidates the old one immediately).

Features beyond the basics

• Catalogue upload: Products → Upload catalogue accepts a text-based PDF, CSV or TXT (Product | Price | Unit, one per line), shows a preview, and adds up to 10 products per upload. A sample file is available in the modal (public/sample-catalogue.pdf).

• Product pictures: when a product is added (upload, form or MCP), a picture is looked up by name from the free Open Food Facts and Open Products Facts databases (no API key). A match requires every word of the product name, so wrong pictures are avoided. Coverage is best-effort: if nothing is found, the offer image shows a letter placeholder and the Products page has a "Find picture" button.

• Offer image: on approval, a 1080x1080 shareable graphic (headline, discount, product pictures, prices, validity) is rendered server-side at /o/<slug>/image with Satori. No AI. The public page uses it as the link preview (og:image).

• WhatsApp sharing: the dashboard's Share image + message uses the device share sheet with the image attached (choose WhatsApp, then search any contact). The message is the share_message that Claude wrote in create_offer (editable before approval). Fallbacks: download/copy image, copy message, per-contact wa.me chat links. WhatsApp does not allow attaching an image through a plain link, hence the share sheet.

Guardrails

Deterministic rules in lib/guardrails.ts (no AI) run on every MCP tool, every dashboard form, and again when you click Approve:

• Only FMCG: blocks adult content, illegal drugs, weapons, hate/extremism, gambling/crypto/loans and prescription medicine terms in product names, groups, offers and messages.

• No personal data in public offers: phone numbers, emails and links are rejected, so buyer details cannot leak into an offer.

• Sanity limits: discount at most 90%, valid price range, no past or far-future dates, max 20 offers per 24h, 30 pending offers, 500 products, 500 contacts per group.

• Offers can only reference the distributor's own catalogue and groups, and are always saved as pending until a human approves them.

• The MCP server also sends the connected model a scope statement (instructions) telling it to refuse non-FMCG requests and never put personal data in offers.

Limitation: the app cannot control what a distributor's own Claude/ChatGPT says in chat outside these tools. It can refuse to store or publish anything outside its purpose, and a human approves every offer.

Tech stack

Next.js 14 (App Router) + TypeScript, Tailwind, Supabase (Postgres + Auth), @modelcontextprotocol/sdk (stateless Streamable HTTP in a Next.js route), deployable to Vercel.

Run it yourself

Prerequisites: Node 20+, a free Supabase project.

1. Clone and install

   git clone https://github.com/adityakumar0707/schemes-broadcast-mcp.git
   cd schemes-broadcast-mcp
   npm install

2. Create the database. In the Supabase SQL editor, run the contents of supabase/schema.sql. (If you created the database from an earlier version, also run supabase/migrations/002_images_and_share_message.sql.)

3. Configure env vars. Copy .env.example to .env.local and fill in:

   Variable	Where to find it
   NEXT_PUBLIC_SUPABASE_URL	Supabase → Project Settings → API
   NEXT_PUBLIC_SUPABASE_ANON_KEY	the anon / publishable key
   SUPABASE_SERVICE_ROLE_KEY	the service-role / secret key. Server only. Never commit it.
   NEXT_PUBLIC_APP_URL	http://localhost:3000 locally, your deployed URL in production

4. Start

   npm run dev

5. Open http://localhost:3000/signup and create an account. (Signup pre-confirms the email so you can start immediately.)

Connect Claude or ChatGPT

1. Log in and open Connect Claude in the dashboard.

2. Copy the connector URL. It looks like https://<your-domain>/api/mcp?token=<your-token>.

3. In Claude: Settings → Connectors → Add custom connector and paste it. In ChatGPT: Settings → Connectors (developer mode / custom connector) and paste it.

4. Clients that support headers can use https://<your-domain>/api/mcp with Authorization: Bearer <token> instead.

Claude.ai cannot reach localhost. For the full Claude flow, deploy first (below). To test locally, use the MCP Inspector: npx @modelcontextprotocol/inspector, choose Streamable HTTP, and enter your local connector URL.

Treat the token like a password. If it leaks, click Regenerate token.

Deploy to Vercel

1. Fork/import this repo into Vercel.

2. Add the four env vars above. Set NEXT_PUBLIC_APP_URL to your Vercel URL (for example https://your-app.vercel.app).

3. Deploy, sign up on the deployed site, and use the Connect page URL in Claude.

Try this demo conversation in Claude

1. "Add my products: Tata Salt ₹22/kg, Surf Excel ₹180/kg" → appears under Products

2. "Create a group called Faridabad with Ramesh 9876543210" → appears under Groups

3. "10% off Tata Salt for the Faridabad group, valid till Friday" → a pending card appears on the dashboard

4. Open the card, edit a field, click Approve → you get a live link and a WhatsApp share button

5. Open the link in a private window (no login) and tap I'm interested → the count updates on the dashboard

Project layout

app/api/mcp/route.ts        MCP endpoint (token auth, stateless)
lib/mcp/server.ts           the 5 tool definitions
app/dashboard/              offers, approval screen, products, groups, connect
app/o/[slug]/               public offer page (+ app/api/o/[slug]/interest)
app/api/signup/route.ts     creates auth user + distributor row
supabase/schema.sql         tables + RLS policies

Scope and limitations

This is an MVP. Intentionally out of scope: WhatsApp Business API (sharing uses wa.me deep links, so the distributor taps send), Google Contacts import, payments, SMS/email notifications, team accounts, analytics beyond the interest count. The "Contact us" link on the public page is a placeholder (hello@example.com); change it in app/o/[slug]/page.tsx.

License

MIT