Connect Claude

Connect your AI assistant to your website

Ask for a change in plain words and it lands in your draft. It only does what you allow, it never publishes on its own, and you can switch it off at any moment.

What it may do

It starts able to read and to edit a draft

That is the default for every new connection, whichever route you take. Editing happens on the draft, so nothing reaches the public website until a person publishes it.

On from the start

  • Read your business, pages and draft
  • Edit the draft – text, images, sections

Off until you turn it on

  • Publish
  • Generate images (spends credits)
  • Read customer records
  • Update customer records
  • Change settings
  • Connect web addresses
  • Invite people
  • Send customer emails
  • Invoices, offers and services
  • Create new websites or companies

If your website was set up by an agency, they may have narrowed this further — the assistant can never reach past the limits set on your account.

For builders

Claude Code, Codex or Cursor

Three commands. Everything runs on your machine; the only thing that leaves it is the approval you give in the browser.

  1. 1

    Install the skills (optional, recommended)

    The CLI installs versioned skills so the agent knows the safe workflow: work on the draft, report honestly, never publish unasked.
  2. 2

    Pair this machine

    Pairing opens an approval page in your browser. It lists every scope the tool asked for and lets you untick any of them, so the grant can only end up narrower than the request. The token is shown once.
  3. 3

    Point the agent at the site

    The link command picks which site this directory belongs to. From here the agent can read the site, edit the draft, and prepare a publish for a human to confirm.
npx @snabbsajt/cli skills install --agent auto
npx @snabbsajt/cli admin pair
npx @snabbsajt/cli link

Or add the MCP server by hand

Pairing already mints the token; this is the same connection, added directly to Claude Code.

claude mcp add --transport http snabbsajt https://snabbsajt.com/api/mcp \
  --header "Authorization: Bearer sajt_live_..."

Any MCP client that speaks streamable HTTP works the same way: the server address is https://snabbsajt.com/api/mcp and the token goes in an Authorization: Bearer header. Use the MCP configuration of your own client for the exact file format.

For website owners

Connect claude.ai by signing in

Nothing to install and no key to keep safe. Five steps, and you can undo all of them by switching the connection off.

  1. 1

    Open Settings → Advanced → Developers

    It is an optional power-user area. You never need it to use SnabbSajt normally.
  2. 2

    Press Connect and name it

    Something you will recognise later, like “Claude on my laptop”. Then choose whether it covers this website or all of them.
  3. 3

    Leave the permissions as they are

    A new connection starts on view and edit. Everything else is off until you turn it on.
  4. 4

    Add it in Claude and sign in

    In Claude, add a connector with the address https://snabbsajt.com/api/mcp and sign in when asked. There is no key to paste — Claude and ChatGPT connect by signing in.
  5. 5

    Ask for a change

    “Add a section about our new service.” It lands in your draft, private until you publish.
If something does not work

Three common ones

The assistant says it cannot do something.
It only has the permissions you granted. Open the connection and turn on what it needs — publishing, for example, is off until you switch it on.
I lost the key.
Keys are shown once, on purpose. Switch the old connection off and create a new one.
Claude asks me to sign in again.
That is how claude.ai connects — by signing in, not by pasting a key. Access is never widened in the background.

Building sites for clients?

You decide which surfaces each client gets, and an assistant they connect is bound by the same limits.