MAQPNA

Learn

Human-in-the-loop approval for AI agents.

Let agents do the routine work, and make the irreversible steps wait for a person.

Updated 1 November 2026.

FAQ

Common questions

How do I require human approval before an AI agent runs a tool?

Put a control point in front of the tools and write a rule that holds the specific tool and arguments you care about. The call waits until an authorised person approves it, and is denied if nobody does in time.

Will approvals slow my agents down?

Only for the calls you hold. Keep holds narrow, allow routine calls by rule, and let agents that can wait retry later with an approval ID instead of blocking.

Can the person who started the agent approve its actions?

They should not be able to approve its risky calls as an approver. MAQPNA refuses such votes. The exception is deliberate user confirmation through CIBA, where the rule asks that person to confirm an action in their own name on their own device.

What happens when an agent triggers a human-in-the-loop approval?

The tool call is held before it runs and an approval request is created with the tool, the masked arguments and the rule. Approvers are notified, and the agent either waits or retries later with the request ID. If enough approvers approve, that exact call runs once; a deny or a timeout means it never runs. Every step is recorded.

What happens if nobody approves in time?

The request expires and counts as a denial. In MAQPNA the default timeout is 300 seconds, and you can change it.

Hold the calls that matter.

Write one rule, and the gateway holds that call until the right people approve it, then records who did.

More install options
curl -fsSL https://maqpna.com/install.sh | sh