Power Platform Community Forum Thread Details

I want to enforce the user confirmation before running any tool of an MCP server. However, I haven't found a way to provide in the adaptive card the tool name that the agent wants to call. This renders this feature basically useless.

How can I do that?

What I want to include in the question:

I want to see that the agent plans on using the tool "plan_..." (and ideally the arguments that it plans on using

You're hitting a known gap. The OOB "Require confirmation" toggle on an MCP tool / connector action renders a generic Adaptive Card and does not expose the planned tool name or arguments as bindable variables, so you can't customize that card to show what the agent is about to do. Voting on it under aka.ms/CopilotStudioIdeas is worth doing.

The workaround that actually works today is to bypass the OOB confirmation and build your own confirmation step inside a topic:

Turn Require confirmation OFF on the tool.
Author a topic that triggers on the user intent that should call the tool (or invoke the topic from the orchestrator via "Use generative AI to fill inputs").
Before the Call action / tool node, add an Adaptive Card (Send a message > Adaptive Card) that you author yourself. Inside that card, reference the variables you have already collected as inputs (e.g. Topic.argName) and hard-code the tool's display name in the card title. You know exactly which tool the topic is about to call, so you fully control the card.
Add a Question node with Multiple choice (Yes / No) below the card.
Wire the Yes branch to the Call action node, No to a polite cancel message.

Why this works: inside a topic, the planned tool name and its inputs are deterministic (you authored the flow), so you don't need a "System.Tool.PlannedCall" variable, you already have the data. The OOB confirmation is built for the autonomous/orchestrator path where the model picks the tool at runtime, but the Adaptive Card schema there isn't templated against the planned call.

Two extra notes:

If you need this for the orchestrator-driven path (model picks the tool, not a topic), the only lever today is Tool description + Input descriptions. The orchestrator will narrate "I'm going to call X with Y" in chat if you instruct it to in the agent instructions ("Always summarize the tool you intend to run and the arguments before invoking, and ask the user to confirm"). Prompt-based, not deterministic.
For audit/observability of which tools fired with which args, hook the agent to Application Insights (Settings > Advanced > Application Insights). The Bot Framework activity payload there gives you tool name + arguments per turn.