Voice-triggered agents become risky when a natural-language request can move directly into a consequential tool call. VoiceGuard MCP demonstrates a transparent checkpoint for that boundary.
In the simulation, a low-value allowlisted request can proceed immediately. A high-value request is paused behind a one-time explicit confirmation challenge. A non-allowlisted action, an over-budget action, or a transcript matching a documented injection phrase is blocked. Every outcome is added to a SHA-256 hash-linked audit trail so a reviewer can detect later modification.
The browser is not a mock disconnected from the backend. It creates MCP client calls that travel through the same Streamable HTTP /mcp endpoint an external host would use. The project uses the official TypeScript SDK and MCP 2026-07-28.
The deliberate scope is a simulated Alexa+ experience: no production customer account, audio stream, biometric signal, payment API, or device integration is claimed. The goal is to make the confirmation and audit semantics easy to inspect, test, and discuss before connecting them to real capabilities.
Built With
- alexa
- mcp
- node.js
- typescript
- zod
Log in or sign up for Devpost to join the conversation.