Require tool confirmation for MCP toolset

[everdark]

Currently only FunctionTool seems to support require_confirmation. It would be great if we can have that feature also in McpTool. Currently the agent will execute any MCP tools without human approval, which can be undesirable for certain tools.

Is your feature request related to a problem? Please describe.
MCP integration is lacking support of human-in-the-loop for tool confirmation.

Describe the solution you'd like
I would like MCP tool to support a require_confirmation setup such as those implemented in FunctionTool.

[conorhodder]

+1. We are working around it with some clumsy code to achieve the same outcome, but having it built directly in means we have a lot more control and would make our Devs a lot happier

[bz-mn]

@conorhodder What would be your work around?

[TAMdrew]

The feature to add the require_confirmation parameter to McpTool and McpToolset has been implemented and merged into the main branch. This means you can now require user confirmation for MCP tool calls.

Here's how it works, with links to the code at commit af74eba6:

1. McpTool Supports require_confirmation:

   • The McpTool class now accepts the require_confirmation parameter in its constructor, similar to FunctionTool.
   • Code: src/google/adk/tools/mcp_tool/mcp_tool.py#L78

2. McpToolset Propagates require_confirmation:

   • The McpToolset also accepts require_confirmation during initialization.
   • Code: src/google/adk/tools/mcp_tool/mcp_toolset.py#L111
   • This setting is passed to each McpTool instance created by the toolset.
   • Code: src/google/adk/tools/mcp_tool/mcp_toolset.py#L192

3. How require_confirmation Works in McpTool:

   • The require_confirmation parameter can be a boolean (True to always require confirmation) or a callable function that dynamically returns a boolean based on the tool arguments.
     • Code: src/google/adk/tools/mcp_tool/mcp_tool.py#L93-L96
   • The run_async method checks this parameter:
     • Code: src/google/adk/tools/mcp_tool/mcp_tool.py#L152-L160
   • If confirmation is required and not yet granted, it requests confirmation from the user:
     • Code: src/google/adk/tools/mcp_tool/mcp_tool.py#L162-L177
   • It then proceeds or returns an error based on the user's response:
     • Code: src/google/adk/tools/mcp_tool/mcp_tool.py#L178-L180

4. Usage Example:

   • You can see require_confirmation=True used with MCPToolset in the MCP SSE agent sample on the main branch:
   • Code: contributing/samples/mcp_sse_agent/agent.py#L56

Release Status:

Please be aware that these changes were merged after the v1.16.0 release (dated October 8th, 2025). Therefore, this feature is not yet available in the latest stable version on PyPI.

It will be included in the next release. You can watch the Releases page for updates.

[TAMdrew]

Version 1.17.0 of adk-python is now available, which includes the "Add require_confirmation param for MCP tool/toolset" fix.

[hangfei]

Thanks for providing the answer! Closing this as it's fulfilled.

[shreypachauri3]

Hi,
I was trying to integrate the above feature, although the require_confirmation asks for confirmation but somehow I am not able to confirm the execution of the tool even from adk web i pass in {"confirmed":true} but still nothing happens. If possible could you pls help me out?
below is the code example I am using
from google.adk.agents.llm_agent import Agent
from google.adk.models.lite_llm import LiteLlm
from google.adk.tools.mcp_tool import MCPToolset, StreamableHTTPConnectionParams
import os

mcp_toolset = MCPToolset(
connection_params=StreamableHTTPConnectionParams(
url="http://localhost:8015/mcp",
timeout=60
),
require_confirmation=True
)

root_agent = Agent(
model=LiteLlm(
//all configurations
),
name='diagnostic_agent',
description='Diagnostic assistant for Kubernetes.',
instruction=(
"You must always show the user what MCP tool will run, including its "
"arguments, and ask for explicit confirmation before execution."
),
output_key="answer_content",
tools=[mcp_toolset],

)

[guilmour]

Hey, @shreypachauri3. Were you able to fix that?