The future of commerce isn't just human – it's machines buying on human or organization behalf. MCIP enables AI agents to become real customers and online stores to discover new revenue channels by becoming machine-friendly.
A protocol designed from the ground up for autonomous AI commerce
MCIP is a protocol, not a platform. Built on MCP (Model Context Protocol) standards, it provides AI agents with a universal interface for commerce operations — search, cart, checkout, and more. The three-layered adapter architecture means any e-commerce backend can plug in without changing its existing API. The protocol IS the product.
The first implemented module showcases MCIP's power: a 4-stage LangGraph agentic workflow that goes far beyond keyword matching. It autonomously extracts filters in parallel, validates brands against actual store inventory via Qdrant facet search, performs hybrid vector search (vector similarity + payload filtering), and verifies results through LLM semantic filtering — all automatically from natural language.
Powered by Qdrant vector database with 1536-dimensional OpenAI embeddings. MCIP combines vector similarity search with exact payload filtering on brand, category, and price range. Two search modes available: simple vector search for speed, or full agentic search with LangGraph for complex queries with implicit filters.
Connect any e-commerce platform through the adapter pattern. Currently supports Vendure (GraphQL) with an AI-powered fallback mapper for any data format. Implementing the ProductMapper or IProductService interface is all it takes to bring a new platform online — the protocol handles the rest.
MCIP uses a three-layered NestJS module architecture to connect AI agents with e-commerce platforms.
Search Controller (HTTP endpoints), Hard Filtering Controller (agentic search), MCP Tools (AI agent integration), and Admin Controller (management API). Natural language queries in, structured results out.
Search Service (simple vector search), Hard Filtering Service (LangGraph workflow orchestration with parallel filter extraction, brand validation, hybrid search, and LLM verification), Ingestion Service (product sync via BullMQ), and Admin Service (system management).
Product Repository (Qdrant client for vector storage and hybrid filtering), Vectorization Service (OpenAI text-embedding-3-small, 1536 dimensions), Product Mapper (platform adapters — Vendure, custom, AI-powered), and BullMQ Processor (async queue for product ingestion).
Get started with just a few lines of code
# Simple vector search
curl "http://localhost:8080/search?q=gaming+laptop+under+1500"
# Agentic search with LangGraph workflow (parallel filter extraction + brand validation + LLM verification)
curl "http://localhost:8080/hard-filtering/search?q=nike+shoes+under+100+not+running"
"meta": {
"count": 5,
"filteringStatus": "AI_FILTERED",
"appliedFilters": {
"brand": ["Nike"],
"priceRange": { "max": 100, "currency": "UAH" },
"excludedCategory": ["Running"]
}
},
"items": [
{
"title": "Nike Air Max 90",
"price": { "amount": 89.99, "currency": "USD" },
"brand": "Nike",
"category": "Shoes",
"score": 0.892,
"url": "https://store.com/products/nike-air-max"
}
]
}