Algolia MCP Supermarket Assistant

This is a submission for the Algolia MCP Server Challenge

What I Built

A shopping bot that gets it. Say "I'm making tacos" and it finds tortillas, beef, cheese automatically. No keyword searching.

Built with Node.js + OpenAI Agents + Algolia MCP. Conversation memory, smart result limiting.

Demo

Repo: campsjos/algolia-mcp-supermarket

curl -X POST http://localhost:4242/api/chat -H "Content-Type: application/json" \
  -d '{"prompt": "I want to make chocolate cake"}'

Auto-searches when you mention food/recipes/weather. No explicit commands needed.

OpenAI Agents call the Algolia MCP server automatically when users mention food/recipes/weather.

Key parts:

• Proactive triggers: Recipe → ingredients, weather → appropriate products
• Result extraction: Parse agent _generatedItems to get Algolia hits
• Smart limiting: 1 result per search if multiple, 4 if single
• Session memory: Conversation context across requests

The integration is seamless - users don't know they're hitting Algolia.

Key Takeaways

Main challenge: AI is unpredictable. Sometimes "I want pasta" would just give recipes instead of searching for ingredients. Other times it would search for random stuff I didn't expect. Spent days tweaking the system prompt until I found the right balance of being explicit about when to use MCP tools without making responses feel robotic.

What I learned:

• Prompt engineering is crucial - Generic prompts don't work. Had to be very specific about trigger words and search behavior
• Agent state parsing is messy - OpenAI Agents structure responses differently each time. Needed multiple fallback strategies to reliably extract Algolia results
• User experience matters - Raw search results overwhelm people. Smart limiting (1 per search vs 4) made conversations feel natural
• Debugging AI is different - Added comprehensive logging because you can't just console.log your way through unpredictable AI behavior

Development process:
Started simple with basic chat, then added MCP integration. Hit the wall with inconsistent AI behavior - sometimes it would search, sometimes not. Realized I needed to treat the AI like a junior developer: give it very clear instructions about what to do and when. The breakthrough was crafting a system prompt that made proactive searching feel natural rather than forced.

Biggest surprise: How seamless MCP servers make tool integration feel once you get the prompting right. The Algolia MCP server just works - all the complexity was in getting the AI to use it consistently.

First time with MCP servers - they make AI tool integration feel natural.