Loading Gatelion
GatelionDocs
Open the console

What your customers see

This page follows one of your customers from connecting their AI assistant to confirming a booking it made. Use it to test your setup and to explain the steps on your own help pages.

The examples use a flight-booking portal called Skyline Air. Your customers see the name you entered as Name your customers know you by on Customer access.

Connecting their agent

The customer, or the company that makes their assistant, adds your Gateway's MCP address to an MCP client that supports MCP sign-in (OAuth). With Claude Code:

Terminal
claude mcp add --transport http skyline <the MCP address>

The MCP address is the one shown for your Gateway on Customer access. It ends in /g/<gateway id>/mcp.

Then, in Claude Code, they run /mcp, choose the server and choose Authenticate. Claude Code opens a browser for the sign-in.

No token or header is needed: the MCP client finds out where to sign in from the Gateway, registers itself with Gatelion the first time, and keeps the sign-in fresh afterwards. If the agent isn't used for 30 days, the customer has to sign in again.

If you chose Wait for your approval for new agents and haven't approved this one yet, the customer sees Agent waiting for approval: Skyline Air has not approved Claude Code yet. Try again once it has. If you blocked it, they see Agent not allowed.

Your login

The browser goes to your login page, the one your customers already use. They sign in as usual, with your passwords, second factors and rules. Gatelion never sees their password; it only learns who they are from your login.

They need to finish signing in and choosing on the consent page within 10 minutes, in the same browser, or start again from their agent.

The consent page

After signing in, the customer sees a page from Gatelion that asks: Allow Claude Code to use Skyline Air for you? It shows:

  • the account they signed in with;
  • the agent's name, and the website it says it's from, with a note that the agent chose these itself and they are not checked;
  • where the browser goes back to if they allow it, as a host name. That host is what really decides who gets the permission, so a customer can spot an agent pretending to be another;
  • the tools the agent will be able to use, by name: only those that are turned on, reached through this Gateway's connections, and that your policy could let this agent run. It lists up to 30, then "and N more", or says you haven't made any actions available to agents yet;
  • that your rules apply to everything it does, including tools added later, that some actions will ask for their confirmation first, and the address of their own page, where they can withdraw the permission at any time.

They choose Allow or Don't allow. With Allow, they're sent back to their agent, which can now call tools for them. The permission stands until it's withdrawn. If the agent ever has to sign in again, the customer sees your login and this page again.

Confirming an action

When the agent asks for something your policy marks for the customer's confirmation, such as a booking, it gets a link (approval_url) and the message Waiting for the customer: ask them to open approval_url and confirm. A well-behaved agent shows the link to the customer.

  1. The customer opens the link. It says nothing about the request until they sign in.
  2. They sign in again with your login. Gatelion asks for a fresh sign-in, even if they're already signed in, so that whoever confirms is the customer, now.
  3. Gatelion checks it's the right customer. Only the customer the agent is acting for can confirm. Anyone else, including your own staff, sees Different account and can't go on.
  4. They review the request. The page says which agent asks to do what with Skyline Air, shows a readable summary, and under Everything Claude Code sent lists every field and value the agent sent. Nothing is hidden. It also shows when the request expires.
  5. They choose Confirm or Don't do it. With Confirm, the Gateway runs the call once, with exactly those values. With Don't do it, nothing happens.

After confirming, the page says the agent can go ahead. If your policy also needs one of your approvers (both), it says Skyline Air also reviews the request before it goes ahead. Either way, the agent learns the outcome by checking the status of the call.

A link works for one decision. Once the request has been decided, has expired or was cancelled, the link shows Link no longer valid before any sign-in, and says nothing else about the request. If it closes while the customer is signing in, the page after sign-in says why, such as Already confirmed, Declined or Expired. If the agent's permission was withdrawn, the page says Permission withdrawn and the request can't go ahead. With both, a customer who opens the link again after confirming sees Already answered.

If you change something that means the request can no longer run, for example you save a new policy, or change or switch off the tool, the request is cancelled and nothing will happen. A customer who is signed in on the page at that moment sees No longer valid: the service changed what agents can do, and they can ask the agent again if they still want it. While the Gateway is stopped, the link shows Service not available.

Withdrawing permission

Customers withdraw an agent's permission themselves on their own page, https://<your Control Plane>/oauth/access/<your organization id>. The exact address is on Customer access; link to it from your site, for example from your account settings. The consent page shows it too.

  1. They open the page and sign in with your login.
  2. They see the agents they've allowed to use your service for them, each with the name (and website, if any) the agent gave itself and the date they allowed it. If they never allowed one, the page says so.
  3. They click Withdraw permission next to an agent. The page confirms that the agent can no longer use your service for them.

The page works for about 10 minutes in that browser; after that, they open it again and sign in. The agent can't renew its sign-in, and its calls for that customer stop after the Gateway's next sync, usually within 15 seconds. Customers only ever see and withdraw their own permissions; withdrawing is recorded in your audit trail as done by the customer. If they want the agent back, they connect it again and choose Allow.

You can still withdraw a permission for a customer who asks: find them under Customers who let an agent act for them on Customer access and choose Withdraw.

An agent that properly signs out, revoking its sign-in with Gatelion, also ends the permission.

Related: Set up customer access · Calling tools

Still stuck? Contact support from the console.