Skip to main content

Require approval for a tool call

Mark a tool as requiring approval and the gateway pauses any call to it until a person approves or rejects the exact call, with its exact arguments. The gate runs at the gateway itself, not in a client-side hook, so it applies to every MCP client connecting to the Virtual MCP, including ones with no hook support at all.

This is a different control from the manual-approval update policy, which holds back a new tool from appearing on a server, and from Agent Monitor's ask-user rule action, which pauses a call inside a supported coding agent's own hooks. Tool call approval gates a specific, already-published tool at the gateway, for any client.

What happens on a gated call​

When anyone calls an approval-required tool, the call doesn't run. The gateway creates a pending approval and returns an error result telling the calling agent to show the user a link. The person opens the link, signs in to MintMCP, and lands on an Approve tool call page showing:

  • The tool name and the Virtual MCP it came from
  • The actual call arguments, as JSON
  • Approve and Reject buttons, with an optional reason field

Once approved, the agent retries the call and it goes through.

An approval is single-use and bound to the exact call: approving one call doesn't approve a different call with different arguments, and doesn't approve a repeat of the same call. The agent has to retry after approval, and the retry consumes it.

Approval requests expire, from 30 seconds to 24 hours, one hour by default. An expired request can't be approved; the caller has to retry to generate a new one.

Approval runs after rules and pre-call middleware, so a call a rule would already block never reaches the approval gate and never burns an approval. The arguments an approver sees are the effective, post-middleware arguments that would actually execute.

Turn on approval for a tool​

  1. Go to Manage store in the sidebar.
  2. Click the MCP server you want to configure.
  3. Select the Tools tab.
  4. Find the tool in its row, and click the shield icon next to the gear (edit) and play (test) icons. The tooltip reads Require approval before execution.

Once it's on, the tool's row shows a violet Approval required badge, and the tooltip changes to Approval required (click to disable).

The toggle either keeps the tool's existing timeout or defaults to one hour. It's hidden without Can Edit access on the connector, so the tool list stays visible without being editable.

Set a custom timeout​

The UI toggle is the on/off switch; a non-default timeout (anywhere in the 30 second to 24 hour range) is set through the Admin MCP's set_tool_enabled tool with a setApproval change, which needs Can Edit access on both the MCP and the connector. Ask your assistant connected to the Admin MCP, for example:

"Require approval on the export_all_data tool on the Analytics server, with a 4 hour timeout."

Both paths write the same per-tool approval config, and the gateway enforces it identically regardless of which one you used.

Audit trail​

Every state change is audited: requested, granted, denied, and consumed, each with the tool, the requester, the approver, and the reason. These export to your SIEM alongside the rest of your gateway traffic; see SIEM export. The stored argument preview follows your log retention setting and is cleared once that window passes.

Next steps​

  • Tool customization: the Tools tab where the approval toggle lives, alongside enable/disable and renaming
  • Admin MCP: the set_tool_enabled tool for setting a custom approval timeout
  • Agent Monitor rules: the ask-user action, a complementary client-side approval step for supported coding agents
  • SIEM export: exporting approval events alongside gateway and Agent Monitor traffic