Skip to main content
Version: 4.0 preview

Connect your own AI assistant

Version
4.0 preview
Updated
View markdown

Audience: store owners, shop managers and the developers who help them · Scope: giving an AI assistant of your own access to the store, and deciding what it may do · Last reviewed: 2026-08-25

What this does — The AI product assistant on your configurator pages talks to visitors. This is the other direction: an AI assistant that works for you — Claude Code, Claude Desktop, or any tool that speaks the Model Context Protocol — gets a door into the store and a set of tools it can use through it. It can then build and tidy products, write the texts and rules, check on orders, tell you what visitors keep asking for, or keep the installation healthy — whichever of those you allow.

The assistant never sees your database directly and never bypasses the store's own checks: every change it makes goes through the same validation a person saving a form goes through, and it can only reach the areas its token was given.

What you need​

  • A store on a web address the assistant's computer can reach. For an assistant running on the same server as the store, a developer can connect it locally instead.
  • An API token for the assistant — see Manage API tokens. The token is what decides what the assistant may touch, so choose it deliberately (next section).
  • An AI tool that can connect to an MCP server over the web, with a bearer token. Claude Code does; so do a growing number of desktop assistants and editors. Its own documentation says where to paste an address and a token.

Give it exactly the access the job needs​

Open API Tokens in the CBX menu, click Add, name the token after the assistant, and use a Start from button:

  • author — the assistant builds and maintains your catalog: products, pages, questions, answers, texts, rules and prices. It cannot see a single customer or order. This is the one to start with.
  • insights — the assistant reads your catalog and the visitor conversations, and tells you what people ask that your store does not answer. It changes nothing.
  • fulfilment — the assistant works orders and can see who they belong to; nothing else.
  • operator — store configuration and installation housekeeping; no customer or catalog access.

Then adjust the boxes if the job is different, and save. Copy the token from the screen right away — it is shown once.

Two boxes deserve a moment's thought before you tick them:

  • customers and orders are personal data. Give an assistant those rows only when its job is about customers or orders, and prefer read unless it really has to change them.
  • code lets the assistant write price calculations as program code — which is program code running on your server. No Start-from button ticks it. Tick it yourself, only for an assistant a developer is supervising, and untick it again when that work is done.

An expiry date is a good habit for an assistant used for one project: the access ends by itself.

Connect it​

Every assistant asks for the same two things:

  • The address: your store's web address followed by /cb-api/v1/mcp — for example https://www.example.com/cb-api/v1/mcp.
  • The token, sent as a bearer credential. Most tools have a field for it; in Claude Code a developer adds the store with one command, passing the address and the token.

Once connected, the assistant sees a list of tools — only the ones its token allows — and starts by asking the store to describe itself: which entities exist, which fields they have, which languages the store runs, and what this store has been customized with. You do not have to explain any of that; it learns it from the store.

What it can do, and what it asks first​

With an author token the assistant can create a product from a brief, add its pages, questions and answers in every language your store runs, write the detail texts the visitor-facing assistant later answers from, set up rules and price calculations, and check its work by reading it back. With insights it can read every conversation the visitor-facing assistant had and tell you which questions went unanswered.

Some tools are marked as destructive — deleting a product and everything under it, importing a product package, running database housekeeping. An assistant that respects those marks will ask you before using one. The store enforces the token regardless: a tool the token does not allow is refused even if the assistant tries it by name, and the refusal says which permission was missing.

Keep an eye on it​

  • API Tokens shows when each token was last used. An assistant token that has gone quiet can be revoked.
  • Every tool the assistant uses is recorded in the store's authorization log with the token's name — a developer can show you the trail.
  • If a token leaks, revoke it and mint a new one; the assistant reconnects with the replacement.

Good to know​

  • The same token works for the store's ordinary API too — one credential, one meaning.
  • The assistant's changes land immediately and there is no undo, exactly as if a person had saved the form. Start it on a staging copy of the store if you want to see how it works first.
  • Nothing here affects the visitor-facing AI product assistant; the two are configured separately and can run together.
Last updated on