MCP server

The MCP server connects AI tools to your Elasticr data. Through it, Claude, ChatGPT or any other AI client that supports the MCP protocol can browse the catalog and prepare changes to categories, products and parameters. Nothing changes until you approve the change.

Experimental feature

The MCP server is still experimental. Its features, their behavior and the way to connect may still change.

About the feature

MCP (Model Context Protocol) is an open standard that AI tools use to talk to other systems. Elasticr uses it to offer the AI tool a set of functions, such as "find products", "show the category tree" or "prepare moving a category". The AI picks which function to call based on what you ask in the conversation. You just write what you need, for example "find products that are in no category and suggest where they belong".

Elly, the assistant in the administration, uses the same functions. She calls them directly inside Elasticr, while your own AI client connects to them over the network. Both follow the same permissions and the same approval of changes.

This page is for the shop's administrator or IT person who wants to connect an AI tool to Elasticr data. Change plans prepared by the AI are listed in the administration here:

Path in administration

AI > AI agent > Plány změn

Switching the MCP server on

Connecting from outside is switched off by default on every installation. Until it is switched on, the MCP server address answers as if it did not exist and the AI client cannot connect. Ask Elasticr support to switch it on.

This switch does not affect Elly in the administration. It only applies to tools that connect from outside.

A user and a token for the AI tool

The AI client signs in to Elasticr with one user's token. To set it up:

1. Create a dedicated user for the AI tool. Everything the AI does runs under this user's name, so the change history shows at once what came through the AI.
2. Switch on API přístup (API access) for the user and save. The token is generated on save and shown in the Token field.
3. Give the user a group with the permissions described below.

How the token is handled and what happens when API access is switched off is described on the Users page (in Czech for now).

Warning

The token works like a password. Anyone who knows it works in Elasticr with that user's permissions. Only enter it in the AI client's settings, never into the conversation itself.

Connecting the AI client

In the AI client's settings, add a remote MCP server with these details:

Setting	Value
Address	https://<your admin address>/platform/mcp
Transport	Streamable HTTP
Token header	X-AUTH-TOKEN, value = the user's token

Elasticr does not offer sign-in through an account (OAuth). The client must therefore be able to send a custom header with the token. If your client cannot do that, it cannot connect.

User permissions

What the AI may do is set by the permissions of the user's groups (in Czech for now):

• API access: without it the server refuses the connection. A user who wants to use Elly needs the same permission.
• Catalog: browsing categories needs viewing the navigation, editing categories also needs editing it, and creating new categories also needs creating it.
• Products: changing product types and parameter values needs editing products. For the AI to create a new parameter value, the user also needs editing product types.
• Parameters and product types: browsing needs viewing product types, changes need editing them, and creating a new parameter or product type also needs creating them.
• Change plans: viewing them to see the plan list in the administration, and editing them to discard a plan.

A function the user has no permission for is not offered to the AI at all. The AI does not know about it and cannot try to use it.

What the AI can do

Shop overview. The AI finds out which user it works as, which sales channels you have, the content language and its limits.

Browsing and analysing the catalog:

• the category tree and category details, including the rule, filters and product count;
• product types, parameters and their values;
• finding products by text, code, type or category;
• explaining why a product is or is not in a category;
• trying out a category rule before setting it;
• finding problems in the category tree, such as categories with no rule or no products;
• products that are in no visible category.

Changing categories: creating, editing the title, address and SEO, setting the rule and filters, moving in the tree, hiding or archiving. Categories cannot be deleted; the AI archives them instead.

Bulk changes to products: adding or removing product types, setting parameter values and creating a new parameter value, such as a new brand.

Editing parameters and product types: creating and editing a parameter, renaming and reordering its values, merging duplicate values into one, creating a product type and editing its parameters. The only thing deleted is a value merged into another; products that had it get the value that was kept. The AI cannot change parameters maintained by an import.

Background job status. Large changes run in the background and the AI keeps track of how far they have got.

The AI works in the shop's default content language. If you have several sales channels, the AI chooses the one the request is about. If it chooses none, it uses the default channel and says so.

How changes are approved

No change to categories, products or parameters happens straight away. The AI first prepares a change plan with a preview of what will change, and only an approved plan is applied.

• Your own AI client: the AI shows you the preview in the conversation and applies it only after you agree.
• Elly: the plan appears as a card in the conversation, and you apply or discard it with the buttons on the card. Elly cannot apply a plan herself.

Every plan:

• is valid for 24 hours, then it expires and can no longer be applied;
• can only be seen and applied by the user who created it;
• is applied in full or not at all. If the affected data changed since the plan was prepared, the plan is refused with the status Zastaralý (Outdated) and the AI has to prepare a new one.

The change plan list in the administration

The Plány změn (Change plans) list shows the plans the AI prepared: summary, area (Katalog, Produkty, Parametry), status, number of operations, who created the plan and when. For each plan you can open its Operace (Operations) and Historie (History). A plan in the status Čeká na schválení (Waiting for approval) can be cancelled with Zahodit (Discard). You cannot apply a plan from the list: that only happens in the conversation where it was made.

Plan statuses: Čeká na schválení (Waiting for approval), Vypršel (Expired), Ve frontě (Queued), Provádí se (Applying), Proveden (Applied), Selhal (Failed), Zahozen (Discarded), Zastaralý (Outdated).

FAQ

The AI client cannot connect, the server "does not exist". What now?

The MCP server is most likely not switched on for your installation. Ask Elasticr support to switch it on. Also check that the address ends with /platform/mcp.

The connection is refused because of permissions. Why?

The user whose token the client uses has no API access. Add it in the user's group.

Can the AI change anything without my consent?

No. Changes to categories, products and parameters are always prepared as a change plan first and applied only after you agree.

The AI says the plan cannot be applied. Why?

Either the plan expired (it is valid for 24 hours), or the data it touches changed in the meantime. Ask the AI to prepare a new plan.

Related documentation

• Users (in Czech for now): API access and the token
• User groups (in Czech for now): permissions
• Integrations overview (in Czech for now): other ways to connect Elasticr