Search projects, hackathons, sponsors and tech

    ← All projects

    Vouch Protocol

    Providers stake capped guarantees on AI-agent work; a confidential test pays clients when it fails

    Demo

    About

    Vouch is a confidential, post-completion outcome-assurance protocol for AI-agent work. Acceptance tests are point-in-time. An AI agent's output can pass the public checks, get accepted, and get paid for — then be revealed as defective by a private or later-run test. Once the fee is released, the client has no recourse and the provider bears no consequence. Vouch closes that gap. A provider voluntarily attaches a provider-funded, capped performance guarantee to a job. During a coverage window after payment, a confidential test runs inside a TEE and can prove a covered failure — automatically triggering a predefined service-credit payout to the client. The verdict becomes part of the provider's onchain performance history. Vouch is not a marketplace, not insurance, and not ordinary escrow. Escrow releases on acceptance; Vouch's novelty is the post-acceptance coverage window plus a confidential trigger that keeps private tests and criteria from ever leaking. Guarantees, payouts, and reputation are authoritative onchain; secrets live only inside the enclave.

    How it's made

    Vouch is a pnpm + Turborepo monorepo. The dashboard is Next.js 16 with Tailwind v4 and shadcn/ui, backed by a NestJS 11 API that serves offchain operational data only — never money, reputation, or secrets. An autonomous agent quotes risk and monitors jobs but never settles funds. Authoritative financial state lives in a Foundry contract (AssuranceHub) deployed to Arc, where the capped guarantee, coverage window, and USDC payout logic run. We integrate Circle's Agent Stack for the agent wallet and USDC movement, giving the agent a real onchain identity that can transact within strict guardrails. Reputation is derived, not stored: a The Graph subgraph built from scratch indexes Vouch's onchain events into an authoritative performance history, which the agent consumes over live GraphQL to make risk-quotation decisions — closing the loop from settlement back into pricing. The confidential verification is a Chainlink CRE Confidential Workflow. The private test and criteria run inside a TEE via handlerInTee; the DON-signed enclave verdict is the sole authorization gate on the guarantee payout, so evaluation logic and repo credentials never touch Postgres, public onchain metadata, the subgraph, or IPFS. Shared Zod schemas, chain constants, and ABIs keep contracts, agent, and API type-safe across the workspace; Prisma 7 backs the rebuildable offchain cache.