NorthCinder MCP Server: A Shopping Agent That Asks Before It Buys

• Makes a checklist for researching a product or seller

  Returns a fixed checklist and limits on searches and source reads for one product or seller request. The tool itself does not browse. Your assistant follows the checklist with its own tools.

  Try: “Make a research plan for the Brooks Ghost 16 in size 10”

  Read · create_research_plan

• Finds products across your stores, ranked by your criteria

  Searches the store connections you set up (Shopify, eBay, Etsy, read-only Amazon) and ranks offers only by price, spec match, delivery, availability, merchant trust and ethics. Sponsored listings are labelled and sit below organic ones, and each result says why it ranked where it did.

  Try: “Find black wool running shoes under $130 and tell me why the winner ranked first”

  Network · search_products

• Adds product facts your assistant found on the web

  Takes product facts your assistant collected with its own browser or search tools and feeds them into the same ranking. The description tells the assistant not to send cookies, raw pages, screenshots or passwords.

  Try: “Add these three offers I found on a small shop to the comparison”

  Write · submit_browser_observations

• Attaches research notes to specific offers

  Links product and seller research to exact offers from one search so each offer shows how ready it is to compare. It never changes the rank or allows a purchase.

  Try: “Attach the seller research to the second offer”

  Write · submit_decision_evidence

• Shows the summary of a past search again

  Re-prints the buyer's brief for a search from this session, with up to three role-based picks. More finalists, rejections and store coverage are in the expanded view.

  Try: “Show me the brief for that search again”

  Read · get_buyers_brief

• Reads your saved shopping preferences

  Returns the preferences saved on your computer, such as sizes, budget defaults, brand rules and delivery defaults. Each entry says whether you stated it or it was inferred from feedback.

  Try: “What shopping preferences do you have for me?”

  Files · get_profile

• Saves or deletes shopping preferences

  Adds preferences you said in the conversation or deletes entries by id, whether stated or inferred. The assistant is told to relay only your exact words and never to guess.

  Try: “Remember that I only buy fair-trade brands”

  Delete · update_profile

• Records your reaction to a result

  Logs a reaction such as not interested or more like this. A reasoned reaction to a brand can create a pending proposal, and a wrong-interpretation note records a correction without learning a preference.

  Try: “Mark the second result as not interested because of the brand”

  Write · record_feedback

• Confirms or dismisses a suggested preference

  A suggestion does not become a preference until you confirm it or a second matching offer backs it. Confirming makes it a stated entry and dismissing removes it.

  Try: “Dismiss the proposal to deny that brand”

  Write · review_preference_proposal

• Sets a price alert

  Creates a standing watch that notifies you when the price reaches a target, through the notification channel set up on your computer. The description says a watch can never buy anything.

  Try: “Watch this jacket and tell me when it drops under $120”

  Write · create_watch

• Lists your price alerts

  Shows every watch, including cancelled and expired ones, with target price, last check and expiry. The notification topic is never shown.

  Try: “List my price watches”

  Files · list_watches

• Stops a price alert

  Cancels one watch by id. The record is kept as voided so the history stays auditable.

  Try: “Cancel watch w_3”

  Write · cancel_watch

• Rates how trustworthy a seller looks

  Returns trusted, known, unknown or flagged for a merchant, with the evidence. Unknown merchants come back as unknown with a reason, not as safe.

  Try: “How trustworthy is this merchant?”

  Network · get_trust_signal

• Starts a purchase with a spending cap

  Step 1 of 3. Creates a pending authorization for one offer with a hard cap. A one-time confirmation code is written to a local file and is not returned to the assistant. Nothing is bought yet.

  Try: “Start a purchase for the first offer with a cap of $130”

  Write · request_purchase_authorization

• Approves a purchase with your code

  Step 2 of 3. You approve by giving the one-time code from the local file, exactly as typed. The description says never to call it with a guessed or invented code.

  Try: “Approve the purchase with code 481-902”

  Write · approve_purchase

• Cancels a pending purchase

  Voids a pending or approved but unused authorization and destroys its code. No code is needed, so saying no is never blocked.

  Try: “Decline that purchase”

  Write · decline_purchase

• Carries out the approved purchase

  Step 3 of 3. Checks the signed approval for the exact offer, quantity, merchant, cap and expiry, and burns its single-use nonce. It then either pays through an agent commerce rail with a payment token or gives you a cart link to finish yourself.

  Try: “Complete checkout for the approved purchase”

  Write · complete_checkout

• Lists your orders

  Shows orders made through NorthCinder plus orders recovered from confirmation, shipping and return emails that you drop in a local folder. It rescans that folder first, and emails are parsed by fixed rules, not sent to a model.

  Try: “List my open orders”

  Files · list_orders

• Records how an order turned out

  Saves outcomes you confirm and reminder dates for warranty or maintenance. Outcomes are never guessed, and reminders only notify.

  Try: “Record that order o_7 arrived and fits”

  Write · record_order_outcome

• Shows one order and its shipments

  Returns the items, linked shipments with tracking history and the return-window deadline. It also writes or refreshes a calendar file for that deadline in the local config folder.

  Try: “Show order o_7 and its tracking”

  Write · get_order

• Adds an order it could not find

  Records an order placed elsewhere or without an email, using only the fields you give it.

  Try: “Add my order from the local shop, $59 on October 3”

  Write · import_order

Watch out for

• It can complete purchases

  complete_checkout can pay through a payment token or hand you a cart link. Every checkout needs a fresh approval for one offer and one unit, with a spending cap and a one-time code. Set a low cap and read each summary before you give the code. github.com

• Your assistant may be able to read the approval code

  The code is written to a local file, not returned over MCP. The README says an AI app with file access as the same user may be able to read it. Do not give the assistant that access, or use a separate OS account. github.com

• It writes and reads local files

  It keeps an audit log, profile, watches and orders in your config folder. It also reads order emails you drop in a mail folder and writes a calendar file for return deadlines. github.com

• Research results are provisional

  The README says no host and model pair is qualified for routine research use. Check the identity, sources and conflicts yourself before you rely on a result. github.com

• Browser results are not store-verified

  Facts your assistant collects by browsing are labelled as reported by the agent. Purchases and unattended watches are not automated until a native store connection confirms the offer. unpkg.com

Security advisories