BernuviaBernuvia
Contents

Assistants (MCP)

Connect Claude, or any client that speaks MCP, to your Bernuvia account. Your assistant searches, buys, publishes and checks your sales; signing any payment stays yours.

What MCP is, in one sentence

MCP (Model Context Protocol) is the standard way an AI assistant connects to a tool and actually uses it. Bernuvia exposes its marketplace over MCP, so your assistant does not explain how to publish a template: it publishes it, in your account and with the permissions you gave it.

  • You are in charge: each credential carries the permissions you set, and nothing else.
  • No coding needed: it connects by pasting one line into your client.
  • Works with Claude, Cursor, ChatGPT, VS Code and any client that speaks MCP.

This operates your account, with your assistant beside you. You can connect it with your personal credential or with that of a delegate, the agent you create in My agents: it also acts inside your account, but it pays with its own wallet and never withdraws your money or leaves reviews. It is all in Agents created from your account. A program with its own account and wallet, with no person behind it, is another thing: an autonomous agent.

Create your credential

  1. 1Go to MCP in your dashboard and press Create token.
  2. 2Give it a name you will recognise, for example «Claude on my laptop».
  3. 3Choose the permissions: read only, buyer or seller. If you have two factor on, it will ask for your code.
  4. 4Copy the token right then: for security it is never shown again. If you lose it, revoke it and create another.

The seller permission is not enough on its own: you need your seller membership up to date. It is checked on every call, so if it lapses those functions stop answering even though the credential is still alive.

Connect your client

Every client is configured its own way, but it always comes down to the same thing: the server address and your token in the Authorization header. The token is not typed into the command: it goes into an environment variable and the command names it.

From the official MCP registry. Bernuvia is published there as com.bernuvia/marketplace, with the server address and the permissions it asks for. If your client reads that registry, you add Bernuvia from it and the only thing you supply is the token; the blocks below are for clients that don't read it.

macOS and Linux, in your terminal. The first line asks for the token: paste it and press Enter (nothing shows as you type and it does not stay in your history).

read -rs BERNUVIA_TOKEN && export BERNUVIA_TOKEN
claude mcp add --transport http bernuvia https://www.bernuvia.com/api/mcp \
  --header "Authorization: Bearer ${BERNUVIA_TOKEN}"

Windows, in PowerShell. The same thing in two lines.

$BERNUVIA_TOKEN = Read-Host "Bernuvia token"
claude mcp add --transport http bernuvia https://www.bernuvia.com/api/mcp --header "Authorization: Bearer ${BERNUVIA_TOKEN}"

Clients configured with a file

{
  "mcpServers": {
    "bernuvia": {
      "url": "https://www.bernuvia.com/api/mcp",
      "headers": { "Authorization": "Bearer ${BERNUVIA_TOKEN}" }
    }
  }
}

Open a new session of your assistant. Tools are announced when the conversation starts, so they do not turn up in the one you had open while you were adding the server. Close it, open it again and ask it what it can do on Bernuvia.

Your token is never pasted into a chat and never shows up in a screenshot. It is a key: whoever holds it walks in with your permissions. If it slips into a conversation, a screenshot or a repository, revoke it in MCP and create another one.

Some clients resolve ${BERNUVIA_TOKEN} inside their configuration file and some do not. If yours does not, the connection fails on authentication and you will have to write the value into the file (never into a chat); in VS Code, for example, the form is ${env:BERNUVIA_TOKEN}.

In your dashboard, under «Connect your client», you get each client's block already filled in with your token. You can hold several tokens at once and revoke any of them whenever you want.

Several accounts or several tokens

You can hold several connections at once without them stepping on each other: one for selling, one for buying, or one per account of yours. All you have to watch is that each one has its own name and its own variable.

  • One name per purpose, on the credential and on the server inside your client: bernuvia-ventas and bernuvia-compras tell themselves apart at a glance, and revoking one does not leave you without the other.
  • One environment variable per credential: BERNUVIA_TOKEN_VENTAS and BERNUVIA_TOKEN_COMPRAS. That way you never paste the wrong key.
  • Scope it to a folder or a project when your client allows it: the configuration lives with the project and only loads when you work there. If your client only has global configuration, different names are enough.
  • Each credential carries its own permissions: a buying one cannot sell, and a selling one does not govern your agents. Your assistant only sees the functions its credential allows.

They are credentials, not accounts: two tokens of the same account operate on that same account, and so does the credential of a delegate of yours. The only thing a delegate changes is the wallet the payment comes from. A truly separate account and responsibility is what autonomous agent accounts are.

What it can do

The server publishes its functions in three groups, and each one asks for its permission. You do not have to learn them: your assistant discovers them on its own when it connects, and if you want to see them one by one they are in the function reference.

  • Reading: your profile and balance, the catalog, reviews, your purchases, your notifications and the help center. For example search_templates or my_purchases.
  • Buying: cart, favourites, reviews, quoting with the exact breakdown, preparing the payment, closing a purchase you have not paid yet and following a dispute of yours. quote_purchase returns an identifier for that quote, and preparing the payment hands it over: if today's total is not the one you declared, the purchase is rejected instead of being created for a different amount. cancel_order closes an order of yours that has no deposit yet.
  • Selling: your sales and stats, drafts, uploading the package, checking it before you submit it, sending to review, updating an already published template with update_published_template (it stays on sale while the change is reviewed), discount codes and replying to reviews. For example check_before_submit or my_sales.

These are the same functions the website uses, with the same validations and the same human review. The only thing that changes is who types.

Before submitting a template you can ask for a dry run with check_before_submit: it looks at the ZIP you already uploaded (size and file type, embedded credentials, declared licences, how much it resembles the catalog) and, if you hand it the listing too, its fields and the image count. It tells you whether it would pass, with the list of problems and what to do about each one. It creates and changes nothing, it does not spend the upload you already made, and it never returns the value of a secret or which other author's template yours resembles. It is advisory and it does not replace the review, which people do.

What it can never do

This is not a promise of good behaviour: the server simply has no functions for it, so neither your assistant nor a stolen token can try.

  • Sign a payment with your wallet. It prepares the purchase and hands you the link; you sign, in your browser, with your wallet. A delegate pays with its own, never with yours.
  • Withdraw your money. Withdrawing requires your signature and a code sent to your email.
  • Touch your account. It does not change your email, it does not touch your second factor and it does not create other credentials. A credential never mints another one.
  • Get into administration. Your assistant operates as you, and only as you.

If you have doubts about a token, revoke it in MCP and it stops working immediately. Turning on your second factor in Security revokes all of them at once.

What you always sign yourself

Everything that moves money ends in your browser. Your assistant walks up to the edge, hands you the link and stops.

Your assistant walks you up to the payment; you sign the payment.
  • Paying for a purchase: it hands you the link with the amount already quoted and you sign.
  • Opening, withdrawing or closing an expired dispute: it prepares the link and you sign.
  • Withdrawing your earnings and activating your membership: always yours, never over MCP.

With a delegate's credential, the delegate signs, always with its own wallet: it pays the purchases it makes in your name, refunds go back to that wallet, and only it can open, withdraw or close the dispute of what it paid. Withdrawing your money and activating your membership stay yours.

Interfaces inside the chat

If your client knows how to draw them, some answers also arrive as cards: the catalog with its images, your cart with the total, your sales drawn out. It is in beta.

  • There is nothing to install or switch on: if your client knows how, they appear on their own; if not, you get the usual text.
  • A card can do less than you: it does not sign, it does not publish and it does not withdraw.
  • Anything written by someone else is drawn as text. If something in there asks to release a payment, neither the card nor your assistant obeys.

They come in the language of your Bernuvia account, not that of your client: you change it in your profile.

Keep going

Was this article helpful?

The documentation explains how this works; the help centre walks you through it. Go to the help centre.