github.com/scadam/retail-mcp · MCP Server · Retail · Front-line

Costa Coffee Frontline AI Agent

An MCP server that powers a Costa Coffee frontline AI assistant for baristas, shift managers, store managers and regional managers. Built with FastMCP and rendered through Jinja2 HTML widgets — but the agent is more than the server. As a Copilot agent it also has native SharePoint, web-search and image understanding, so it combines live data, documents and photos in a single conversational flow no mobile app can replicate.

View on GitHub ↗ See the widgets

19Tool modules

17Interactive widgets

5Stores modelled

15Demo scenes

Why an AI agent — not just a mobile app

A mobile app gives you menus and screens. This agent gives you a thinking partner that reasons across sources.

Weather + footfall

Turns signals into staffing advice

Combines the forecast, local events and store history into specific, actionable guidance — "rain + rail chaos = lots of people sheltering with a coffee; double-check oat milk and cups."

Image understanding

Reads a photo and acts

A barista uploads a photo of the milk fridge at 8°C — the agent flags the food-safety breach, cross-references the SharePoint manual, and logs the incident in one conversational turn.

Travel disruption

Knows who's affected by name

A rail strike surfaces the specific staff whose commute is hit, and recommends cover actions — not just a generic alert.

Multi-step workflows

Chains work in one conversation

Stock → order → log action → notify manager. Every interactive element uses the OpenAI Apps SDK sendMessage API to send follow-on prompts back to Copilot.

The widget surface

Every widget is Costa-branded, renders in a mobile-sized Copilot panel, and turns each tap into a guided follow-on prompt.

Daily sales dashboard widget — **Daily Dashboard**Sales KPIs, hourly revenue chart and top products.

Shift rota widget — **Shift Rota**Weekly grid with crew avatars and "Find Cover" prompts.

Stock levels widget — **Stock Levels**Health doughnut, category icons and one-click stock orders.

Recipe card widget — **Recipe Card**Drink image, clickable allergen cells and barista tips.

Training module catalogue widget — **Training Catalogue**All 12 modules with search and REQUIRED badges.

Training video player widget — **Training Video Player**Progress tracker with guided Copilot actions.

Customer feedback widget — **Customer Feedback**NPS, star ratings and a 12-week trend chart.

Regional benchmarks widget — **Regional Benchmarks**Cross-store performance league table.

Sam's Wednesday morning

It's 7:15am on a rainy Wednesday and there's a rail disruption on the line. Sam unlocks the store and, hands-free, asks: "Morning. What do I need to know about today?" — the agent runs the weather and travel tools in parallel, spots that two crew members commute from Woking, and warns of a commuter surge from 7:30am.

Over the next hours it reads a photo of an out-of-range milk fridge and logs the incident, flags oat milk running critical and drafts an emergency transfer, updates the rota to cover a late arrival, coaches a new starter through a Caramel Latte allergen question, and drafts a reply to a queue-complaint tweet — all in one continuous conversation. That's the difference between a screen and a thinking partner.