Whose account it runs on
For a skill that signs in to an app, Connection sets who agents act as when they call it.
The choice applies to every agent the item reaches, to Omni and to MCP clients at once, and takes effect from the next call. A skill whose actions need a setting per connection cannot run on each user’s sign-in; the menu says Not available.
When you add an app from the library, its Connect panel starts on Each user signs in (No secret is saved; everyone uses their own account). To run it on one account instead, use Or connect one shared account below it and sign in. The shared sign-in is stored in Secrets under the name in Saved as secret, which follows the connection name until you edit it.

Connecting an app: each user signs in by default, with one shared account as the alternative. Shown with sample data.
Allow API keys
For an app people connect with their own account, Allow API keys decides whether they may paste their own API key or token instead of signing in (Connect with your own key). On by default. Off keeps the app on sign-in only: Use a token disappears from My tools, the chat sign-in card and agents’ skills, and AI tools get only a sign-in link. Apps whose provider takes its own sign-in only, such as Google, Microsoft and Atlassian apps, have no switch.
Allow API keys on an app's card in the Skills registry. Shown with sample data.
Which actions may run
Tool access limits what the item may do: All tools, Read-only tools (actions that only read) or Custom set. Tools added later decides what happens to actions the skill gains later: On, Read-only or Off. An action outside the allowed set is refused everywhere with a message that an admin turned it off.
Ask for approval
Ask for approval > Choose tools picks the actions whose every call waits for someone to approve it: All tools (including actions the skill adds later) or a list. The item then shows Approval required, and the summary names who approves. The organization’s ask applies on top of anything an agent’s owner set, and a project cannot lift it.
In an agent or Omni, a held call is an ordinary approval: a card in the conversation and a request on the approvers’ Settings > Approvals page (Approvals). In an MCP client, the call answers that it waits for an approval, and the request reaches the approvers named for the person and the client (for example “Claude Code”). Once it is approved, the person asks their client to run the call again with the same arguments. A declined call does not run.
A call made without an agent, through the invoke API or xpander_invoke_connector_operation over MCP, is held the same way. It answers 202 with "status": "awaiting_approval" and an approval_request_id; once someone approves, repeat it with the same arguments. A declined call answers 403. A call on a key that names no person, such as the organization’s default key, is refused, since nobody can be asked.
Who approves
The first match decides:- The people and groups named under Who approves on the item’s
Ask for approvalcard. - The organization’s default: the Who approves button at the top of the Skills registry. It decides every held call whose item names nobody.
- With nobody named in either place, the organization’s admins.
Adding an app from an AI client
Admins working in an MCP client connected to the gateway can add an app from your organization’s catalog without opening xpander: the client offersxp_create_connection, which takes the app’s name (for example Gmail). A name that matches several apps returns the choices. The new connection is set to Each user signs in, appears in the Skills registry, and the admin gets their own sign-in link in the answer. Asking for an app the organization already connected says so; giving a new name adds a second connection under that name. An app that signs in with a key, not each person’s account, is added from the Skills registry instead. Nothing is added when the app’s catalog entry cannot be used, and the answer says why.
Other members do not see xp_create_connection; they ask an admin to add the app.
