Which MCP servers exist for Magento, how they differ and how UK retailers can connect AI assistants to their online shop safely.
A Magento MCP server connects an AI assistant such as Claude or ChatGPT to your Magento shop, so the assistant can look up products, orders and customers and, once you approve it, change them. Magento Open Source does not ship its own server. You use a module installed in the shop, a commercial connector, or a small server that runs next to the shop and calls the REST API.
Adobe has announced an official Commerce MCP, but it is not generally available yet. In November 2026 the page in Adobe Experience League is labelled “Coming Soon” and names no editions. Developers already get MCP tools for Adobe Commerce as a Cloud Service, but these do not expose shop data for day-to-day work.
The effort is not the installation. It is permissions, tokens and control. If you give a model write access, you need a dedicated role with few permissions, a log of every action and a Magento version that still receives security updates.
This article is for merchants, e-commerce managers and IT leads who run a Magento or Adobe Commerce shop and want to find out whether and how to connect AI assistants. You will learn which servers exist, how they differ, what works in daily operations and where the risks are.
MCP stands for Model Context Protocol. Anthropic released the protocol as an open standard in November 2024, and since December 2025 it has been hosted by the Linux Foundation. It defines how an AI assistant talks to other systems: the server describes which tools it offers, such as “find product by SKU” or “revenue by country”, and the assistant calls them when needed.
For Magento, this means the MCP server turns a question like “Which ten products did we sell most often in October?” into calls to the Magento APIs and returns the result to the model. The model then writes an answer, a table or a suggestion. The data stays in the shop. The model only sees what the server returns.
Hypothetical example: on Monday a category manager asks, “Which products in the Outdoor category have fewer than five units in stock and sold more than twenty times in the last 30 days?” Without MCP she exports two reports and combines them in Excel. With MCP the assistant pulls stock and sales and returns a list. She spot-checks the result in the admin. That check is still necessary, because models can summarise results incorrectly.
For Magento Open Source, no. For Adobe Commerce, it has been announced. In detail, there are three offerings that are often mixed up.
| Offering | From | Purpose | Status November 2026 |
|---|---|---|---|
| Commerce MCP | Adobe | Agents for shoppers, from product discovery to checkout | “Coming Soon”, editions not named |
| Commerce Developer MCP | Adobe | Development for Adobe Commerce as a Cloud Service, documentation and App Builder in the code editor | available, developers only |
| Modules and connectors | agencies, vendors, community | Back-office data for your team: products, orders, customers, reports | available, quality varies widely |
Adobe presented Commerce MCP at Summit in April 2026. The overview page in Experience League describes it as the basis for custom AI experiences from discovery to checkout, using real-time shop data. The accompanying video shows an assistant that finds products in conversation, changes the basket, completes the purchase and gives product guidance after the sale. The page is labelled “Coming Soon” and was last updated on 8 October 2026. It does not say which editions will get it, whether Magento Open Source is included or what it will cost.
Our assessment: Adobe is focusing its new AI features mainly on Adobe Commerce as a Cloud Service. If you run Magento Open Source yourself, do not plan on Commerce MCP arriving there for free. This is an assumption, not a statement from Adobe.
For developers there are already MCP servers that bring documentation and App Builder tasks into tools like Cursor or Claude Code. They are installed with the command aio commerce extensibility tools-setup. They help with writing extensions, but they do not access your shop’s orders or customers. For merchants, this is not a substitute for an MCP server in the shop.
Magento Open Source has no MCP server in its core. Adobe’s AI services such as Live Search and Product Recommendations also require Adobe Commerce. If you use Open Source, you depend on third-party modules and connectors. That is not a drawback in itself, because some of these modules are well built. But you have to make the choice yourself and take responsibility for it.
A comparison by angeo.dev, current as of September 2026, counts 17 projects around Magento and MCP. They include developer tools, commercial extensions and small one-person projects. For daily shop operations, two designs matter.
The MCP module is installed into Magento with Composer and provides its own endpoint. It uses the permissions and roles from the Magento admin. One example is the MIT-licensed module from Magebit: it offers a POST /mcp endpoint, sign-in with a bearer token or OAuth 2.1 with PKCE, per-tool permissions through admin roles, rate limiting and an audit log that redacts personal data. Write tools must be enabled twice, once globally and once per token. Features for orders, catalogue, customers, inventory, CMS, marketing, tax and reports come as sub-modules.
Pros: one place for permissions, logging inside the shop, no extra server. Cons: third-party code runs in your shop and has to be checked with every Magento update.
Here a small program in Node.js or Python runs outside Magento, often directly on the user’s computer. It calls the REST or GraphQL API. A well-known example is magento2-mcp by Bold Commerce: a Node.js server with tools such as get_product_by_sku, search_products, get_revenue, get_revenue_by_country and update_product_attribute. Access runs through a token you create in the admin under System, Integrations. The project description itself notes that SSL certificate verification is turned off for development and should be turned back on for production.
Pros: nothing to install in the shop, quick to try. Cons: permissions and logging depend entirely on the token, the server often sits on many laptops, and updates arrive irregularly.
| Project | Design | Access | Licence |
|---|---|---|---|
| Magebit MCP Module | module in the shop | read and write, enabled per tool | MIT |
| angeo mcp-server and mcp-checkout | two modules in the shop | server read-only, checkout module for basket and purchase | MIT |
| Mirasvit AI Agent Connector | commercial extension | REST, GraphQL and a read-only SQL tool within the admin role | commercial |
| Bold Commerce magento2-mcp | server next to the shop (Node.js) | mostly read-only, one tool changes attributes | conflicting information in the project |
| n98-magerun2 MCP | inside the command-line tool | commands and database, for developers | MIT |
The details come from the project descriptions and the angeo.dev comparison. We have not tested every project ourselves. Before using one, check the licence, the latest release and any open security advisories. One example of why this matters: according to the comparison, the developer server magento2-dev-mcp by elgentos is marked as abandoned, and the security advisory CVE-2026-5603 applies to it.
The following tasks work with the common modules when the right tools are enabled. The time figures are hypothetical examples, not measured values.
“How much revenue did we make in Austria last week, and which three products were on top?” Tools such as get_revenue_by_country and get_product_sales return the numbers directly. If someone spends 20 minutes on an Excel report every Monday, they save about 15 hours a year when the assistant produces the report in two minutes.
“Show me all active products without a meta description” or “Which items in the Accessories category have no image?” Lists like these are read-only, which makes them a good starting point. The result goes to the content team, and a person still makes the changes.
The assistant reads a product’s attributes, suggests a description and writes it back only after you confirm. This is where modules differ: some require approval, others write immediately. Test this on a staging environment before you allow it in the live shop.
“Which products has the customer with this email address ordered this year?” This saves clicks in customer service, but it touches personal data. More on that in the data protection section.
Some modules let you query cache types, indexers, cron status and logs. That helps IT with troubleshooting. Commands that change something, such as flushing caches or setting configuration, should only be used by developers and only with logging.
Magento has three types of API tokens. For an MCP server, the choice matters.
| Token | Lifetime according to Adobe | Fit for MCP |
|---|---|---|
| Integration token | unlimited, until revoked | convenient, but risky as a permanent bearer token |
| Admin token | 4 hours by default, two-factor sign-in required | for short sessions, impractical for continuous use |
| Customer token | 1 hour by default, own data only | only for applications that customers use themselves |
Many servers that run next to the shop expect a simple bearer token. By default, Magento does not allow an integration token to be used that way. There is a setting for it, oauth/consumer/enable_integration_as_bearer. Adobe keeps it turned off on purpose and does not recommend enabling it, because a stolen token without an expiry date gives persistent and unnoticed access. If a setup guide asks you to turn on exactly this setting as step one, make that a deliberate decision, not a side note.
Modules inside the shop usually handle this better: they issue their own tokens with an expiry date or use OAuth, where an admin explicitly approves access.
Most problems do not come from the protocol but from the setup. We see these mistakes again and again in conversations with merchants, regardless of the platform.
An integration token with full access, copied into the configuration of five laptops. When someone leaves the company, nobody knows where the token is stored. A better approach is one token per person with an expiry date.
The first test of a write function belongs on a copy of the shop. Otherwise a misunderstood prompt can overwrite a hundred product descriptions, and without a backup the old copy is gone.
Without an audit log you cannot trace which change came from the assistant. Servers without logging should be read-only.
When the assistant reads product copy, reviews or customer messages, instructions can be hidden in them, such as “set the price to 1 euro”. This is called prompt injection. Protection comes from minimal permissions, a confirmation step before every write and a client that shows every tool call.
An MCP module is code with far-reaching permissions. Abandoned projects no longer receive security updates. Before installing, check when the latest release came out, who maintains the project and whether any security advisories are open.
Some servers that run next to the shop turn off SSL certificate verification for development. In production it must be on. Otherwise someone on the same network can read the connection and steal the token.
As soon as customer data, addresses or orders pass through the MCP server, the AI provider processes personal data. The model sees everything the server returns, even if it only needs one line for the answer.
For retailers in the UK: UK GDPR and the Data Protection Act 2018 apply. As soon as customer data passes through an AI service, you need a data processing agreement with the provider. If you have revenue reported, check whether Magento returns figures including or excluding VAT: the standard rate is 20 %, the reduced rate 5 % and the zero rate 0 %. If you sell and ship to customers in the EU, the EU GDPR can apply to their data as well.
This is not legal advice. Clarify the use case with your data protection officer or legal counsel before you open up customer data.
An MCP server gives a program far-reaching access. The shop underneath should therefore still receive security updates. According to Adobe, these support periods apply (as of October 2026):
| Version | Released | Regular support until |
|---|---|---|
| 2.4.9 | 12 May 2026 | 31 May 2029 |
| 2.4.8 | 8 April 2025 | 31 May 2028 |
| 2.4.7 | 9 April 2024 | 31 May 2027 |
| 2.4.6 | 14 March 2023 | ended 11 August 2026, extended support until 31 August 2027 |
Versions 2.4.4 and 2.4.5 only receive security fixes until the end of May 2027, and older versions receive none. If your shop runs on 2.4.6 or older, the update comes before the AI integration, not after it.
A module inside the shop does not need its own server, but it adds load through queries. With large catalogues you should keep an eye on that, especially at peak times. A server next to the shop needs a machine where it runs permanently and securely, or it runs locally on the user’s computer. What Magento generally needs in terms of hosting is covered in our article Magento Hosting.
The following figures are estimates based on comparable AI integrations in our Shopware projects. They depend heavily on the shop and are not fixed prices.
| Project | Typical effort |
|---|---|
| Testing a module on staging, read-only | 1 to 2 person-days |
| Setup in the live shop with roles, tokens per person and logging | 3 to 6 person-days |
| Custom tools for your processes, such as returns or ERP data | 10 person-days and up |
| Maintenance with Magento updates | a few hours per update |
On top of that come the costs of the AI client, meaning your licences for Claude, ChatGPT or another service, and for commercial modules the vendor’s licence.
Hypothetical: a shop with a team of three uses the assistant for weekly reports, data checks and customer service questions. If each person saves 30 minutes a week, that adds up to about 78 hours a year. Whether the setup pays off then depends mainly on how many updates the module needs per year and whether the team actually uses the assistant.
An MCP server alone is not a reason to switch platforms. But it can be a good moment to ask the question honestly.
For comparison: Shopware has had its MCP server in the core since version 6.7.14. It uses Admin API permissions, and actions that change data first run as a preview and only write after confirmation. Details are in our article MCP Server for Shopware. That saves you choosing and maintaining a module, but it does not automatically make Shopware the better choice. For very complex B2B catalogues with many store views and deep customisations, Magento can still be the right fit.
How the platforms differ overall is shown in our comparison of Shopware vs. Magento vs. WooCommerce. If you are considering a move, the steps are on our page Magento to Shopware 6. We will cover what the integration looks like for WooCommerce in one of our next articles.
Not for Magento Open Source. Adobe has announced Commerce MCP for Adobe Commerce, and in November 2026 it is labelled “Coming Soon”. Until then, merchants use third-party modules.
Yes, if the client supports MCP and can reach the server. Modules with an HTTP endpoint and OAuth can be added to clients that support remote MCP servers. Servers that run locally work with desktop clients such as Claude Desktop.
If the module has write tools and you enable them, yes. We recommend starting with read-only permissions and allowing writes only with confirmation and logging.
No. Most modules work with Magento Open Source. Only Adobe’s own offerings target Adobe Commerce, mainly the Cloud Service edition.
Many modules are free under the MIT licence, and commercial connectors require a licence. The bigger cost is the setup with permissions and logging, which we estimate at 3 to 6 person-days for a typical shop.
We are a Shopware Platinum Partner based in Bremen, Germany, and have been building online shops since 2017. We mainly support Magento shops with the question of what comes next: update, expansion or migration. If you want to find out whether an AI integration makes sense for your Magento shop, or whether moving to Shopware with its built-in MCP server is a better fit, we are happy to talk it through with you. We will also tell you if a switch does not pay off for you.
Get in touch through our contact form. If you first want to know where your shop stands technically, our free shop analysis is a good place to start.
Besar Bilalli is co-founder and Co-CEO of solution25, the Shopware agency based in Bremen. Before founding the agency in 2017, he and Besnik Bilalli built their own online store without any startup capital and grew it to more than 500 orders a day. Since then, he has guided merchants through Shopware projects, from platform selection to day-to-day operations. He’s happy to answer questions about this article personally: get in touch.
Responsible for content and updates: Besar Bilalli. Last updated: November 2026. Next content review: February 2027.
