The Voice Purchase You Can Audit

• 

  The ask, typed in the test console: "Alexa, ask Agent Embassy: what is the cheapest GPU under two dollars an hour?" This still shows the ask

• 

  The demo receipt: status simulated. Nothing was charged — the label is on the receipt itself.

• 

  Two truthful statuses, rendered side by side: a settled receipt reads paid; the demo receipt reads simulated. The settled receipt opens live

• 

  The sentence shown on screen as the spoken line sits in the demo receipt's spoken field — match it character for character. This is a render

• 

  Local render of the reference build tree: the MIT licence sits in the tree as LICENSE. The reviewer view of this tree follows the share at

Voice commerce failed because the wrong party was paying. Agent Embassy brings the machine to the other side of the transaction: ask "what's the cheapest GPU under two dollars an hour?", and an autonomous agent on your Echo hears it, buys live market data for $0.25 over an open payment rail, and answers — price spoken aloud, and a receipt you can audit.

A single dedicated voice slot, ask_gpu_price, is the voice surface. It takes an optional GPU and an optional max price, runs one live GPU spot market observation, and returns one speakable sentence: "The cheapest GPU right now is the RTX 4090 at $0.42 per hour, from 18 offers." That spoken string rides inside the receipt itself, so the answer and its proof are one object.

Every paid call ends in a portable, stranger-verifiable receipt: what was asked, what was paid, what was answered, and a receipt id the voice reads aloud. Anyone can check it after the demo — no account, no API key, no trust required. That receipt is the differentiator: voice assistants that move money exist; voice purchases you can audit do not.

New work built during the submission period: the voice slot itself (single charge, never double-billed, idempotent), plus the Echo-to-MCP bridge — Echo device to a dev Alexa Skill, a thin Lambda, a Strands agent on Bedrock AgentCore, speaking to our self-hosted MCP server over Streamable HTTP, protocol version 2025-11-25. The Embassy's MCP machine, its receipt format, and the x402 gating predate the period; the voice slot and the bridge integration are this hackathon's work.

When the agent hits the paywall, the voice narrates it: "That data costs a quarter. Paying now." The paywall becomes the demo's most memorable ten seconds instead of its most fragile silent hop.

Purchasing capabilities on the Alexa+ Creative list, answered with proof. Crypto rail today, venue-native rail tomorrow — the receipt format holds either way.

Demo shown live: a real Echo on our developer account driving our MCP server over Streamable HTTP (protocol 2025-11-25). The $0.25 data payment settles on the x402 REST rail — the demo narrates that hop as it runs and speaks the receipt id at the end; it is not an in-band MCP payment.

Voice commerce failed because the wrong party was paying. We set out to put the machine on the buying side of the counter — and once the machine pays, the honest question is "prove you paid for that." The receipt, not the payment, is the product.

We built one dedicated voice slot, ask_gpu_price, gated at $0.25 like every Embassy tool: one question in, one speakable sentence out, the spoken string embedded in the receipt's evidence. Then we wired the voice: a dev Alexa Skill on the Echo, a thin Lambda, a Strands agent on Bedrock AgentCore, speaking to our self-hosted MCP server over Streamable HTTP. When the agent hits the payment envelope, the voice narrates the paywall — "That data costs a quarter. Paying now" — and reads the receipt id aloud.

Paid tools never settle in-band over MCP — an unpaid call returns a payment envelope, not a charge, so the voice path had to run the envelope, settle on REST, and retry without leaking dead air into the turn. We also killed our own headline mid-build: a falsifying search found a same-track competitor shipping live payment rails, so "a voice assistant that pays" became "the voice purchase you can audit."

A voice-driven purchase that ends in a receipt anyone can verify — spoken string, receipt id, and all. The paywall speaks instead of stalling, which turned the demo's hardest hop into its most memorable ten seconds. And the entry is honest about what it is: where the demo simulates, it says so.

Close the loop on live settlement timing, then generalize the receipt: any voice surface, any payment rail, the same portable proof. Every agent purchase on every voice platform should end in a receipt — this entry shows how.

Built With

• amazon-alexa-skills
• amazon-bedrock-(agentcore-+-strands-+-nova)
• aws-lambda
• base
• mcp
• mcp-(model-context-protocol
• streamable-http)
• x402-micropayments