Skip to main content

Human-in-the-loop operations

Data and scan tools observe; they never change anything. Anything that would modify your system — killing a process, clearing a cache, removing a registry entry, deleting a shortcut — goes through a human approval step. The AI can ask; only you can approve.

How a request flows​

  1. An MCP client calls a request_* tool (e.g. request_kill_process).
  2. The server enqueues an approval request and returns {"status": "pending", "request_id": …} — it does not act.
  3. The app raises a system-tray notification and a status-bar message.
  4. The request appears in the Operation approvals queue on the MCP tab.
  5. You click Approve or Deny.
  6. The client learns the outcome by polling get_operation_status (or list_pending_operations).

The approvals queue​

The MCP tab's Operation approvals (human-in-the-loop) table shows each pending request — the operation kind, its details, and Approve / Deny buttons. Resolved requests move to the History panel with a timestamp, the decision, and the result.

Clicking the tray icon (or its notification) brings the window forward and jumps straight to the MCP tab so you can act quickly.

Mocked execution​

Approving an operation today runs a mock: it records what it would have done (for example MOCK: would have killed PID 1234) and changes nothing on your system. This lets you exercise the full request → approve → result loop safely. The human-in-the-loop machinery is real and complete; wiring the approved action to a real effect is the only remaining step.

caution

Treat every request on its merits. The approval dialog shows exactly what the agent is asking to do — read it before approving, the same way you would review a command before running it.