Article

MCP tools: what a model may call, and the person who can say no

The MCP tools specification expects a human in the loop who can deny a call. What that means for a tool that moves money.

1 min readAgenticOpenFinance editorial teamGenerated automatically

In the Model Context Protocol, a tool is a function a server exposes for the model to call. Each tool has a name, a description and an input schema, and may declare an output schema and annotations that describe its behaviour.

The tools section of the specification is direct about safety. For trust and safety, there should always be a human in the loop with the ability to deny tool invocations. Applications should make clear which tools are exposed to the model, show when a tool is invoked, and present confirmation prompts so that a person stays in the loop. The specification also says clients must treat tool annotations as untrusted unless they come from a trusted server: a tool that calls itself read-only is not proof that it is.

In a finance flow, this suggests a simple design rule. Separate looking from doing: resources and read-only tools for balances and statements, and narrow tools for actions, such as "draft a payment to an existing payee" rather than "pay anyone". Put the confirmation prompt on the action tools, and show the person the amount and payee before the call goes ahead.

Limits. The specification does not mandate a particular user interface, and "should" is not "must". Whether a confirmation actually happens depends on the host application. A bank exposing tools cannot see the host's screen, so it should still enforce its own limits and authentication on the server side.

Sources: Model Context Protocol specification: Tools, Model Context Protocol specification (version 2026-07-28)

#protocols#MCP#human in the loop

Related posts

Hear about new posts and events first

We email you when we announce an event or publish an important guide or report. A few times a month at most.