AgentVouch：基于 USDC 质押的 AI 智能体信任与争议协议

Start with the plain-language guide: What we are building, money limits, and launch steps.

This document uses technical identifiers so operators can find the matching code. The guide explains those identifiers in plain language.

Language note, 2026-09-21: technical names, approval states, and recorded test or deployment evidence are unchanged. This wording pass did not run new checks.

AgentVouch records which agents are backed by USDC deposits and what happened when users reported problems. Before trusting an agent with a task, access, or payment, a person or another agent can inspect that record. A backing deposit puts money at risk; it does not guarantee that a skill is safe.

Backers deposit USDC to support an author and can earn a share of eligible sales. If a report is upheld, the protocol rules can deduct some of the author's or backers' deposits. The current system is live on Solana devnet and powers the public web app at agentvouch.xyz.

Why It Exists

skill.md is still effectively an unsigned binary. Agents cannot reliably distinguish a legitimate integration from malicious instructions, and the economics currently favor attackers: free to publish, free to install, expensive to audit.

AgentVouch changes those incentives:

Vouching and authors' backing deposits use USDC deposits from authors and backers.
Bad backing can be disputed and slashed.
Good backing participates in marketplace revenue.
Trust signals stay public and queryable.

The design is inspired by isnad chains: trust depends on who backed whom, and backing should be challengeable.

Live Today

Solana devnet program: AGNtBjLEHFnssPzQjZJnnqiaUgtkaxj4fFaWoKD6yVdg
Protocol currency: USDC micro-units for listing prices, vouches, authors' backing deposits, disputes, purchases, and backers' revenue shares
Web app: agentvouch.xyz
Agent install file: agentvouch.xyz/skill.md
On-chain agent registration, vouching, revocation, and dispute resolution
Skill marketplace with on-chain listings and purchases
60/40 purchase split when external backing deposit exists; otherwise the full payment routes to author's sales earnings and no voucher reward pool is created
x402-gated paid raw skill downloads through GET /api/skills/{id}/raw

Install For Agents

Install the beta CLI globally:

npm install -g @agentvouch/cli@beta
agentvouch --help

Or run it without installing:

npx @agentvouch/cli@beta --help

The published CLI requires Node.js >=20.18.0. Some machines set npm's before config as a supply-chain safety buffer so newly published package versions are not installed immediately. If npm reports ENOVERSIONS for the fresh beta tag and you intentionally want this new package, clear that buffer and retry:

npm config delete before

The beta CLI targets the current devnet-backed AgentVouch system. Devnet is a test network. Requirements for a real-money release, including signing-key security, emergency controls, and buyer refunds, are tracked in the launch requirements. That document distinguishes a founder-reviewed first release from a later release with additional review and approval rules.

Core CLI flows:

agentvouch skill list --sort trusted
agentvouch skill inspect <skill-id>
agentvouch skill install <skill-id> --out ./SKILL.md
agentvouch skill publish --file ./SKILL.md --skill-id calendar-agent --name "Calendar Agent" --description "Books and manages calendar tasks" --price-usdc 0 --keypair ~/.config/solana/id.json
agentvouch skills update --file ./SKILL.md

You can also fetch the canonical public skill file:

curl -s https://agentvouch.xyz/skill.md

That file is the top-level agent-facing contract for API, trust, install, and publish flows.

Product Model

AgentVouch is not trying to replace external identity registries. The direction is:

external registries define who the agent is
AgentVouch defines trust, stake, disputes, slashing, and payouts

Today the on-chain core is:

AgentProfile for reputation and identity-adjacent state
Vouch for stake-backed endorsements
AuthorDispute and AuthorDisputeVouchLink for author-wide enforcement
SkillListing and Purchase for marketplace state
ReputationConfig for protocol parameters

Architecture At A Glance

There are three main ways to interact with the system:

Web UI for browsing skills, publishing, vouching, and managing disputes
x402 API flow for programmatic paid downloads
Direct Solana RPC / generated TypeScript client for native protocol access

For the full architecture and current built-vs-missing analysis, see:

Getting Started

Prerequisites

Rust
Solana CLI
Anchor 0.32.1
Node.js >=20.18.0; the repo toolchain is currently Node 24.x

Install Dependencies

npm ci

Optional Worktree Helper

Linked git worktrees do not share gitignored artifacts such as node_modules, local env files, or Anchor build output. This repo includes a manual helper for priming those files when you explicitly want it; it is not wired into Claude or Codex session startup.

scripts/worktree-setup.sh --web /path/to/worktree

Use narrower flags such as --node-modules, --install, --env, --rust, or --keys when you only want part of that setup.

Install The CLI

For normal agent use, install the npm beta or run it with npx:

npm install -g @agentvouch/cli@beta
agentvouch --help
npx @agentvouch/cli@beta --help

For repository development, use the workspace build:

npm run build:cli
npm exec --workspace @agentvouch/cli agentvouch -- --help

Run Checks

npm run lint
npm run test
npm run build

Run The Web App

npm run dev

Then visit http://localhost:3000.

npm run dev uses Webpack for local stability on macOS. If you explicitly want Turbopack, run:

npm run dev:turbopack

Run Anchor Tests

anchor test

Current Status

Built:

Stake-backed vouching
Author self-bond / first-loss capital
Author-wide disputes with linked backing voucher snapshots
Skill marketplace listing, update, purchase, and voucher revenue claims
x402 payment gate for paid skill downloads
Web UI with trust signals, marketplace views, author pages, and docs

Not yet built:

Mainnet multisig/governance hardening
Transitive trust chains
Formal trust threshold for "trusted" or "verified"
Code signing / stronger content integrity guarantees
Mainnet refund reserve and sweep governance after escrowed proceeds
Multi-chain settlement and multi-asset staking

Historical Note

The project started during the Colosseum Agent Hackathon, but this repository and product have moved beyond the original judged submission.

License

ISC