One layer for custom integrations and AI

MCP (Model Context Protocol) is an open standard that lets an AI application discover and call tools of another system. The ecommerce MCP server in MitoOps exposes business data — orders, products, stock, shipments, analytics and reports — as tools over a normalised layer that already joins several stores — 213 tools across 60 permission scopes today.

No. Claude.ai connects over OAuth: you give it the MCP address, sign in to MitoOps and choose scopes and stores on the consent screen. A key is only needed by your own program or an agent that calls the server.

Not orders, documents, refunds, shipments or stock quantities: no tools exist for that. Writes concern fourteen things and all sit behind sensitive scopes that are never pre-selected on the consent screen: a disabled automation draft, which only a person in the app can enable, or a revision of a running automation, which only a person loads and publishes (the running version stays unchanged until then); a discount coupon from the parameters you give it — creation never overwrites an existing coupon, explicit update and deletion have a preview and confirmation, deletion also has its own scope, and a coupon with a live dependency is not deleted; likewise with a preview and confirmation it turns on or adjusts the box that shows the price after the code next to products on your website, and the discount rule for campaigns (the store’s maximum discount); and affiliate programme management — approving or suspending a partner account, portal texts and help articles, templates, campaigns, document policy and a partner’s store access; new legal wording arrives only as a draft you publish in the app. Deciding a commission has its own sensitive scope: the assistant may approve, reject or cancel one specific commission you name (rejection with a reason, a paid commission is never reverted); bulk approval and payouts stay in the app only. The fifth is a change to a store’s availability list, such as a new Preorder availability or different delivery hours: the assistant only proposes it with a preview, and a person with the right permissions writes it to the store by approving the proposal in the app. The sixth is a change to the claims settings, meaning processes, statuses, tags and e-mail or document templates: the assistant only proposes it with a preview, and a person with the right permissions writes it by approving it under Claims, Settings. The seventh is a claims return shipping route and the eighth product and category descriptions and SEO texts: the assistant only proposes them with a preview (it cannot change price, stock or code this way) and a person writes them by approving them; if you deliberately switch on autonomous writing for that area, they are written immediately on behalf of the person who granted access and only with their permissions. The ninth is the shop heart, the brand profile the AI tools write by: the assistant may read it and propose a change, which applies only after a person approves it in the AI centre; there is no autonomous writing for it. The tenth is a reply to a review: the assistant prepares a proposal in the store’s language and only a person allowed to reply publishes it under Warehouse, Reviews; autonomously, if you deliberately switch it on, only replies to 4 or 5 star reviews and at most 50 a day; with a preview and confirmation it also turns on the block of reviews from your other stores on the product page, writing nothing to the store. The eleventh is store category management and the product gallery: the assistant shows a preview (before and after), creates a category hidden, shows or hides it, schedules it from–to or deletes only an empty one, and in the gallery adds an image from an address, changes the order or main image or removes an image; it writes only after a person approves, and from outside only if you deliberately switch on Autonomous execution for this area. The twelfth is reports: the assistant writes manual costs (create, edit, retire), the report note and the export layout after a preview and approval, from outside only in Autonomous execution mode; it generates a report whose file is downloaded with a hash, and it does not change automatic advertising costs, orders or documents. The thirteenth is products: the assistant writes text and identity edits, prices, availability, the main category and visibility on the website with a from–to schedule, each behind its own sensitive scope, and creates a product that is always hidden, after a preview and approval, from outside only in Autonomous execution mode for that area; it does not change the stock quantity. The fourteenth is funnels in Analytics (for example a quiz on your website): the assistant creates, edits, pauses or archives a funnel after a preview and approval, from outside only in Autonomous execution mode for Funnels; it changes only the measurement settings in the application, not the store or GA4. The assistant also only proposes availability matching between stores, and only a person switches synchronisation on. The capability kernel, a closed catalogue and the adapters hold the line, and the REST API has no write at all. For each area you choose Manual approval (AI prepares a proposal that is written only after you approve it in the app) or Autonomous execution (AI runs the permitted change right away, on behalf of the person who granted access); every step is in the AI change log.

Only if a person who has that right allows it explicitly: the scopes for reports and purchase prices are sensitive and not pre-selected on the consent screen. Without them, purchase price, gross profit and margin never appear in a response.

Yes. MitoOps is built for several stores, markets and currencies. One connection reads from every store you ticked at consent or when issuing the key.

From the Start plan with one active access key. Growth has two, Pro five, Custom by agreement. The free trial and Beginner do not include the interface.

REST is called by your program, which decides what to read and when. With MCP, the AI model discovers and calls tools. Both paths sit on the same capabilities and return the same business data.

MCP has to be supported by the application the model runs in (the client), not by the model itself. MitoOps provides a standard remote MCP server with OAuth; Claude.ai is the production-verified client. Any other application that supports remote MCP with OAuth connects the same way.

So that your system learns about an event immediately, without polling: a new order, a payment, a status change, a created or delivered shipment, a stock movement, an invoice or a claim. You pick the events and stores per endpoint; failed deliveries are retried and every attempt is visible in the log.

The REST API and MCP are reads on request: your program or AI application asks and gets an answer. A webhook goes the other way: MitoOps sends your system a short signed message that something happened, with identifiers. Your system then reads the detail over the REST API.

Every delivery carries an HMAC-SHA256 signature over the timestamp and body, plus the event ID and timestamp in headers. The endpoint secret is shown once and can be rotated at any time. The receiver verifies the signature with a constant-time comparison and rejects messages older than five minutes — the exact procedure and code samples are in the documentation.