An MCP server that knows what to bundle with what.
Install @marketbasketanalysis/mcp, paste your API key, and your AI agent can answer 'what should I cross-sell with the customer's current cart?' or 'what's a good bundle for an out-of-season item?' against real merchant order data. Basket AI: four named agents (Discovery, Bundle, Insight, Replenishment) over eight typed tools, MCP-protocol-compliant.
(One snippet, one restart)
Why this fit
Most LLMs are great at copy but bad at recommendations: they invent products, pair things that don't actually move together, and have no idea what's in the store's catalog. MBA's MCP server is the missing piece, a tool an agent calls to get real co-purchase data backed by the merchant's own order history. Works with any MCP-compatible host: Claude Desktop, Claude Code, Cursor, Cline, the OpenAI Agent SDK.
The right primitives for the job
Eight typed tools, four named agents
Basket AI groups eight tools (get_recommendations, find_substitutes, get_bundle_for_cart, propose_quote_bundle, score_cross_sell, analyze_basket, predict_reorder, propose_subscription_bundle) into Discovery, Bundle, Insight, and Replenishment agents. Each ships a JSON-schema input contract so the host can validate before calling.
Works with every MCP host
Claude Desktop, Claude Code, Cursor, Cline, OpenAI Agent SDK. One snippet in the host config + one restart.
SSRF-guarded base URL
The MBA_API_BASE env var is validated against an allow-list: rejects loopback, link-local, cloud-metadata IPs. Operators can't accidentally (or maliciously) point the MCP at their own internal network.
Timeout + retry
15s per request, with exponential-backoff retry on 429 + 5xx. Hung upstream doesn't block the host indefinitely; runtime validation rejects malformed responses with sanitized error messages.
Real auth
Bearer API key minted in your MBA admin (Shopify / Magento / WC). Hashed at storage; plaintext only shown once at creation. Revoke per-integration without affecting other hosts.
B2B-aware
propose_quote_bundle is designed for sales-rep agents on Magento B2B: takes a customer ID, returns recommended bundles tuned to that account's purchase history.
What you skip
The friction we're explicitly cutting out.
- Building a custom HTTP client that doesn't know the recommendation API's shape
- Hardcoding product catalogs into your agent's system prompt
- Inventing fake product pairs because the model has no real data to pull from
- Re-implementing auth, rate limiting, and timeout handling per agent project
- Letting agents query internal networks via untrusted MBA_API_BASE values
Install once. Query from anywhere.
The MCP server is free + open. The only paid piece is the MBA backend itself (per-platform pricing on the pricing page). If you're a developer building an agent against a Shopify / Magento / WC store, the MCP server is the most direct path to real recommendation data.
Quick-start install