Features · Security & Backup · MCP Per-Server Toggle
Security & Backup · MCP Per-Server Toggle

Hide a host from the AI,
without unplugging it.

One switch per host: "AI may not see this". MCP tool calls referencing the host return "not found". The toggle itself is web-auth-only — the AI cannot flip it back, by design.

MCP exposure
web-prod-04 visible to AI
db-prod-01 hidden
vault-01 hidden
the toggle itself web auth only
so the AI cannot switch itself on
— Why this matters

The safest thing you can tell an AI about a host is nothing.

01

Read access is still access.

A host holding customer data or key material has no business being described to a model, however well the prompt is written.

02

A switch the AI can flip is not a switch.

The toggle is reachable from the web UI only. It is not exposed as an MCP tool, so nothing on the other side of the gateway can turn itself back on.

03

Hidden means absent, not filtered.

A call naming a hidden host does not come back redacted. It comes back as not found, which is the honest answer to a question that should not have been asked.

— What's included

What the toggle actually does.

One flag per server
On by default. Off means the host stops existing as far as the gateway is concerned.
Enforced in middleware
Any request whose URL names a hidden server is refused before a handler sees it.
Applied to every listing
Servers, status, logs, databases, containers, cron jobs, backups and exports all filter through the same scope — not each one remembering to.
Body lookups too
Creating an incident or running a recipe against a hidden host by UUID fails the same way.
Visible in the UI
A badge on the server detail header, and a column in the list, so nobody has to guess which hosts are exposed.
— Works with

Connected capabilities

— FAQ

Questions, answered honestly.

Can a hidden host still post heartbeats? +
Yes — the toggle only affects MCP visibility. Everything else (alerts, metrics, agent) works normally.
Why web-auth-only and not API-with-elevated-scope? +
Reducing key-exfiltration risk. A leaked API token must not be able to reveal hidden hosts. A leaked human session is a separately-managed risk.
Default state? +
Visible. The toggle is opt-out — most fleets want the AI to see most hosts.

A clear "the AI may not look here" switch.

MCP Per-Server Toggle on every plan with MCP enabled.

Get early access Back to Security & Backup