HazelPiper Creativehazelpipercreative.com

OpenAI WebMCP Challenge

Approval Desk

The agent does the work, the human holds the pen

Approval Desk proves that WebMCP can expose useful agent labor while preserving human authority. The tool list changes with the work. The agent can ask. The agent cannot approve.

An agent has already run on this page. On the left is the queue and the live tool surface. In the middle is whatever is open. On the right is the decision it cannot make for you.

WebMCP leverageDynamic tools, typed schemas, annotations and abortable human gates.

ExecutionLive static app, public repo, no dependencies and a working product surface.

ImpactApprovals, bookings, pricing and publishing are trust states, not just UI states.

CreativityThe website exposes where the agent must stop.

Checking for WebMCP support

On the desk

Registered right now

0tools on this page

Reading the page's own list

    Twelve baseline tools. Thirteen while a human decision is pending. Filled dot changes state. Hollow dot only reads. Gold appears and disappears with the work.

    Judge mode

    Runs the same agent script, then tries to approve by code so the refusal is visible. A person still has to click to close the gate.

    The memorable moment: scripted approval fails, but real approval works.

    Waiting on you

    The gate is real

    An approval settles only on a trusted user gesture. A click created in code is refused, which is what stops an agent approving its own work.

    What happened