The AI assistant is a separate agent application: you state the task in words, and it reads what it needs itself, does the work itself, and reports on what it did. It is tied to the store by a bridge — the MCP server that ships with Melbis Shop and speaks to the engine over the same protocol as the program itself.
Any application will do, as long as it speaks MCP: Claude Code, Goose, Cherry Studio and others. Melbis Shop is not bound to any of them: the application names itself on the first connection and gets all of its instructions there — nothing has to be set up for it in the store in advance.
Three things are required for this to work, and all of them are mandatory:
Do not confuse it with AI Help — that one is built into the Workbench editor, sees only the text you show it, and changes nothing (developer guide, “AI Help”). The rule is simple: fits in a single file — AI Help; needs the store itself — the assistant.
It sees the whole store:
| Area | What it does |
|---|---|
| Project files | reads and edits modules, views, statics, images; keeps the version history, clears the cache |
| Data | reads and changes tables as a pool of steps over a single connection, with locks and cache marking |
| Trees | the catalog, reference directories, and any others: creates, moves, and deletes nodes with index recalculation |
| Item files | attaches product photos and attachments, and fetches them back |
| Store copy | fetches the store’s entire code, profiles, or a database dump — the
same as the program’s “Copy store” button; config.json is
not included in the copy |
| Storefront | opens store pages as a visitor and reads the debugger report |
| Documentation | reads the manuals and table descriptions straight from the distribution — which is why its answers about Melbis are specific rather than general |
“System → Connection → AI Assistant” — and there is nothing to set up there. The tab holds a ready MCP server configuration and a “Copy” button: copy it and paste it into the settings of your agent application, where it asks about MCP servers. Nothing else is needed from you — which store is open the assistant works out by itself.
The same configuration, but tied to a single store, the program puts
into the store’s own folder — the .mcp.json file. An
assistant started from it works with that store only,
whatever is open in the program. This is for applications that open a
folder as a project: open the store’s folder, and the correspondence
about different stores does not pile up in one heap. The file is
rewritten every time the program starts, so adding anything of your own
to it is pointless.
There is no login or password here: the assistant signs in to the store as you, with the same pair you used to sign in to the program on the “Connection Parameters” tab.
It obtains that pair on its own, by one of two routes. The program is running — it answers the assistant straight from its own memory. The program is closed — from the Windows registry, where the password is stored encrypted, but only if the “store passwords” box is ticked. There is no third route: naming the password in correspondence is not allowed — the assistant will not accept it, and it never appears there.
The “Check Connection” button walks the whole path for real — license, password, rights — and explains in words what is wrong, before the first session.
The AI assistant works in your store with your own hands, so the question of what it is allowed to do is settled where it is settled for people: in the user’s rights.
It has no separate account. What it has is a separate branch
in the operations registry — “AI Assistant” — and that is not a
formality: the Load right for a person and the
AI Assistant → Direct access → Development → PHP Scripts → Load
right for their assistant are different rows. So you can keep your own
hands full and leave the assistant with read access only — and the other
way around, whoever has not been granted the branch at all simply has no
assistant.
Three consequences follow, worth keeping in mind:
The assistant does not work around prohibitions: when refused, it names the missing right — and names it as a path through the rights tree, not as an internal command name, because a path is exactly what you see in the program.
The only exception to all of the above is a store that has just been
installed: while the admin.here file lies in its root, the
engine checks neither the license nor the password, and the assistant
works right away. It will warn you about this itself: in such a store
anyone who knows the address becomes an administrator, so finish the
installation before going live.
State the task as a goal, not as a sequence of actions: “the home page needs a bestsellers block” rather than “open such-and-such file and add such-and-such line.” It will work out the arrangement itself and lay out a plan before changing anything — it gives separate warning about editing files and about changing data, and waits for your agreement.
The first task in an unfamiliar store goes slowly: the assistant
looks around — it reads the project map, the documentation, the modules.
This is a preparation stage; after it things go noticeably faster, all
the more so because it writes down what it has understood into
the store’s own memory and next time does not start from
scratch. That memory lives in the database rather than on the computer,
so it travels with the store; the notes belong to the login they were
made under — another employee has their own. The mcp folder
next to the store has nothing to do with this: it holds disposable
material — downloaded pages and images — and is wiped without
regret.
The assistant is not a replacement for a developer. It quickly does what you understand yourself, and its edits are worth reading as a colleague’s edits: before publishing to the live storefront, look them over.
Everything described above is the same in any Melbis store. On top of it the assistant has AI Tools — ready-made commands written for your store: create a product the way it is done here, close an order by your rules. The owner sets them up, and they are handed out one command at a time to each employee — see the next section, “AI Tools”.