FastMCP builders: how are you handling approvals and auditability for tool execution?

[Ram9199]

I'm exploring a question that came up while building MCP tooling:

For tools that can execute shell commands, modify files, access browser state, use credentials, or interact with local services, how are you handling:

• approvals
• delegation
• execution records
• post-run auditability

Specifically, after a tool runs, how do you answer:

• What ran?
• Why was it allowed?
• Was approval required?
• Is there a durable record of the decision?

I've been experimenting with a small evaluation path around:

blocked
-> delegated
-> allowed once
-> blocked again
-> audit verified

using Agent_Sudo:

https://github.com/Kisyntra/Agent_Sudo/blob/main/docs/evaluate_5_minutes.md

I'm not looking for architecture reviews or integration discussions. I'm trying to understand current FastMCP practices for tools that touch shell, files, browser state, credentials, or local services.

Where did you get stuck?