The part of northcinder I would read first is the checkout rule, not the product comparison. According to the README, a recommendation is not permission to buy, and every checkout needs a fresh approval for one exact offer and one unit, signed and usable once. For anyone who has designed authorization for an agent, that reads less like a shopping feature and more like a scoped capability token.
The project describes itself as an open-source MCP server for comparing products and asking the buyer before purchase. Its README opens with a position. It argues that big marketplaces are building shopping agents that search one catalog and steer the buyer toward that platform's checkout, and that this may be convenient but is not independent advice, because the marketplace still decides what can be seen and makes money when the agent closes the sale.
That is the README's framing, and it names no specific marketplace. Its response is software you run alongside your own AI app. The README states that the repository owner does not operate a NorthCinder service, and that there is no NorthCinder account or cloud service.
The "Buying stays a separate decision" section is short, so its mechanics deserve a close read. Every checkout needs a fresh approval for one exact offer and one unit. The signed approval includes the merchant, variant, price, known total, and spending cap. It can be used once.
Read that as an access control design. The approval binds to specific purchase parameters rather than to a session or a general mandate to shop. A second unit or a different variant is outside its scope. The README also says that starting a search or price watch does not give NorthCinder permission to buy anything, so a watch running in the background carries no purchase authority on its own.
Payment handling follows the same restraint. Per the README, NorthCinder rejects raw card details rather than storing or forwarding them. A supported automated checkout can use an opaque payment token. The other path is a cart link handed to the buyer, who finishes the purchase in their own browser.
There is one more gate on the store side. The README says a native connection must confirm the exact offer before checkout or an unattended watch. When a native connection is missing, the AI app can keep researching with its own browser or search tools, but NorthCinder accepts product facts, not cookies, raw pages, passwords, or page instructions. The README states that exclusion without describing how it is enforced.
The README says NorthCinder reruns the ranking locally and writes recommendations, approvals, and checkout attempts to a local audit log. The audit log and purchase approvals stay on your computer. Order outcomes stay local and attach only to the purchase they belong to, and the README says the tool does not silently rewrite the buyer's profile.
The ranking rules are written as plain statements. Seller payment never improves ranking. Sponsored offers stay labeled and below organic results. Unknown seller history remains unknown instead of being guessed safe or unsafe. The README points to separate ranking, trust, neutrality audit, and checkout documents for detail, and it scopes its own checks: they cover the offers NorthCinder received, not the completeness or truth of a store's catalog.
Output is usually limited as well. The README says it normally shows no more than three useful choices, explains why each result ranked where it did, and lets you inspect the other finalists, rejected offers, and facts that could not be verified.
The research flow carries explicit warnings. Before doing research, the MCP host should read the relevant product or seller research guide, call create_research_plan with the actual request and exact subject, then follow the returned checklist with the research tools it already controls. Research can decide whether an offer is ready to compare, but it cannot add ranking points. If sources disagree or do not identify the exact product or seller, the result stays provisional.
The README goes further and states that no host and model combination is currently qualified for routine research use. It tells readers to treat every research result as provisional until the buyer checks its identity, sources, conflicts, and unknowns.
That caveat belongs next to the approval model in any evaluation. The consent gate governs spending. It does not make the upstream research trustworthy, and the README does not claim it does.
You need Node.js 20 or later and an MCP-capable AI app. Running npx northcinder init saves your configuration on your computer and prints the MCP entry for your AI app. Local mode is keyless and runs the MCP server and search engine together in one process, using a temporary loopback port.
If you run the engine separately, the README says to set `NORTHCINDER_API_KEYS` on the service and configure the client with `NORTHCINDER_SERVICE_URL` and the matching `NORTHCINDER_CLIENT_KEY` bearer credential. Non-loopback bearer connections must use HTTPS. Built-in adapters cover Shopify, WooCommerce, eBay, Etsy, and read-only Amazon comparison, and the README says NorthCinder reports when a store was unavailable or not configured.
**GitHub:** [https://github.com/cinderline/northcinder](https://github.com/cinderline/northcinder)
*Curated by [Agent Palisade](https://www.agentpalisade.com) — practical AI for small and mid-sized businesses.*