Killswitches

A killswitch stops one person’s MCP tool calls on a chosen set of servers, right away or on a schedule, until it is lifted or expires. The gateway refuses matching calls before they reach the server, while the person keeps their account, roles, and sessions. That makes it a fast, reversible way to contain a problem without a disciplinary-looking change to the account.

Killswitches are managed on the person’s Access tab under Identity > Identities, so the decision is made while looking at what that person has been doing.

Access requirements

Managing killswitches requires the org:admin scope, which the Admin role includes, and a signed-in dashboard session. API keys cannot manage killswitches.

When to use a killswitch

• Incident response. Pause one person’s tool calls on the affected servers while the incident is investigated.
• Suspected compromise. A flagged laptop or an unusual risk finding calls for stopping activity before the facts are known.
• Runaway AI clients. A client acting as the person keeps calling tools in a loop or against the wrong system. The killswitch holds even if the client reconnects.

A killswitch applies to calls a member makes through their own signed-in MCP sessions. It does not cover API keys or registered agents, even ones the person owns. For a lasting change to who may use a server, edit the server’s Team Access rules instead. For a compromised identity, also disable the account in the identity provider.

Start a killswitch

• On the person’s Access tab, select New killswitch.
• Under Which MCP servers, choose All MCP servers, which also covers servers added later, or Selected servers.
• Under Schedule, choose when it starts (now or later) and when it ends (until lifted or at a set time).
• Write a Public message shown to the member, which appears in the error their client receives, so leave confidential details out. Add an Internal note for other admins.
• Select Review impact to check the scope, the schedule, and any overlapping killswitches, then select Turn off MCP tool calls.

The same form is available from the person’s row menu on MCP sessions and the Members page.

What the person experiences

Each tool call on a server in scope fails with a forbidden error that carries the public message and the code mcp_tool_calls_paused. This applies to hosted, remote, and tunneled servers. Everything else keeps working: the person stays signed in, clients can still connect and list tools, and servers outside the scope are unaffected. If the gateway cannot complete the check, it refuses the call rather than letting it through.

Lift or change a killswitch

Select a killswitch on the person’s Access tab to open its record, which includes a version history of every change and who made it.

• Edit killswitch changes the server scope, schedule, or notes of an active or scheduled killswitch. Servers removed from the scope regain access immediately.
• Lift killswitch ends the restriction. The dialog warns if other killswitches still cover the same person, because lifting one does not lift the others.

A killswitch with an end time expires on its own. Every change is also recorded in the audit logs.

How killswitches differ from other controls

Control	What it stops	Who it affects	Ends sessions
Killswitch	Tool calls on the servers in scope, for a set period	One member	No
Revoke a session	The current connection, though the client can sign in again	One session or client	Yes
Disable a server	All traffic to the server	Everyone	Yes
Suspend an agent	All requests made with the agent’s credentials	One agent	Its credentials stop working

Revoking a session and starting a killswitch are independent. To stop a person and force a clean reconnect, use both.