NorthCinder's MCP server lets AI agents shop only with buyer approval

cinderline/northcinder was created on 17 August 2026 and has reached 1,219 stars and 8 forks in its first 17 days on GitHub. It is a Model Context Protocol server that gives an AI shopping agent a way to compare products across five marketplaces and a mandatory approval step before any purchase.

What the server does

NorthCinder includes built-in adapters for Shopify, WooCommerce, eBay, Etsy, and a read-only Amazon comparison. The agent queries product listings across all five through one MCP interface, and the server returns each result with the source attributed so the model cannot silently prefer one marketplace over another. Every ranking decision comes with a written specification the agent has to state, which the README describes as a transparent ranking rather than a hidden, unexplained recommendation.

The server runs on Node.js 20 or later on the buyer's own machine. The README says NorthCinder never needs the user's AI provider key and does not send searches, settings, or local history back to the repository owner. Raw payment details are refused rather than stored, so a compromise of the local process would not expose a card.

The approval step

The rule that distinguishes NorthCinder from a marketplace-owned agent is the step where the buyer must approve. A purchase is not something the agent can complete on its own: the server presents the ranked options with their reasoning attached, and a human has to confirm before the checkout call is made. The README frames the project against agents that operate inside one catalog and direct the buyer toward that platform's checkout, which is how most vendor shopping agents released now work.

Where the code lives

The project is MIT-licensed. The runtime is JavaScript with a pnpm workspace layout, and the private site workspace requires Node.js 22.12 or later. NorthCinder does not name specific MCP clients it has been tested against; the README states only that any MCP-capable AI application should be able to talk to it.

Agentic commerce is an active product category this year, and every large marketplace is releasing its own agent. The pattern that comes with those agents is the one NorthCinder is designed to replace: the agent runs inside the marketplace, sees only that marketplace's catalog, and completes the sale on the marketplace's own checkout. A team building a buying agent on top of one of those has already picked the seller before the first prompt.

NorthCinder is built on the view that the alternative pattern, a neutral comparison layer with a required human confirmation, is worth having as an open-source component rather than a vendor product. For a team building any workflow that ends in a purchase, whether that is reordering supplies, comparison shopping, or procurement, the interesting questions are whether the ranking is good enough on real catalogues, and whether an audit log of what the agent saw and why it chose one option can hold up in a real dispute later. The approval step is the part that makes that audit possible.

Source

• cinderline/northcinder on GitHub