Connected agents

Connected agents lists the AI assistants that have connected to your tenant over MCP — Claude Desktop, Cursor, and anything else speaking the protocol. One row per connector, with a Revoke button.

It’s the answer to a question that gets more awkward the longer you leave it: which assistants can currently read our data, and who attached them?

What’s on the page

Each connector shows:

• Name — the client that connected, as it identified itself.
• Created — when the connection was authorised.
• Last used — the last time that agent actually called Saiku, or Never if it authorised and then sat idle.

That last column is the useful one. An agent connected four months ago and never used since is someone’s abandoned experiment holding live access to your cubes, and it costs you nothing to revoke it.

Revoking

Revoke invalidates the agent’s access token immediately — the next call it makes fails, and the person using it has to reconnect from scratch. There’s no grace period and no undo.

Nothing else is affected: revoking a connector doesn’t touch your API keys, your cubes, or anyone else’s connectors.

Tip

Revoking is the right reflex, not a last resort. Connectors are free to re-establish — the person reconnects in under a minute — so there’s no reason to leave an unrecognised or unused one in place while you work out whose it was.

How agents get here

This page only manages connections; it doesn’t create them. An agent appears here after someone points their client at your MCP URL and signs in through the OAuth connector flow — see MCP server for the connection reference, both OAuth and Bearer key.

Note

OAuth connectors need Team or above. On a smaller plan, agents connect with a read-only API key instead, and those appear on API keys rather than here.

What an agent can do once connected

Read, and nothing else. An MCP connector gets the same six read-only tools every agent gets — list cubes, describe a cube, search members, run a query, preview a query, drill through — and no way to change a schema, a connection, or a saved query. The full list is on the MCP server page.

Which cubes it can reach is a separate question, decided by the access the connecting user already had. An agent never sees more than the person who attached it.

Related

• MCP server — the protocol, the tools, and how to connect an agent in the first place.
• API keys — the other way agents authenticate, with a cube allowlist you set yourself.
• Audit log — what those agents have actually been querying.