Kanverse ProofStamp demonstrates bounded AI execution. GPT-6 Astra can inspect state and request an action through WebMCP, but execution stops at an explicit human authorization gate inside Roblox. After approval, Roblox independently gathers live Bitcoin and ISS evidence, creates a machine-readable ProofStamp, and returns an authenticated completion receipt that Astra can verify.
GPT-6 Astra orchestrates a bounded real-world-style action workflow through WebMCP. It inspects the ProofStamp relay state, creates exactly one action request, and then stops at an explicit human authorization boundary. After a player authorizes execution inside Roblox, Astra reads the authenticated completion receipt and verifies the resulting Bitcoin block and ISS evidence. Astra coordinates and verifies the workflow but cannot execute the Roblox action itself.
Hey
I built Kanverse ProofStamp to explore a simple but important question: how can AI agents request real actions without silently gaining execution authority?
In this demo, GPT-6 Astra can inspect state and request a ProofStamp through WebMCP, but it must stop at a human authorization boundary inside Roblox.
A real player then presses AUTHORIZE & EXECUTE. Only after that does Roblox fetch live Bitcoin block and ISS evidence, create the ProofStamp, and return a verified completion receipt that Astra can inspect.
AI requests → Human authorizes → System executes → AI verifies.
I’d love feedback on whether this authorization-and-receipt pattern could be useful beyond games for agent tools, apps, devices, and real-world actions.
Kanverse GPU Borrow
Kanverse GPU Borrow