VRYFY

The Trust & Verification Layer for AI Agents

Don't trust. VRYFY.

VRYFY is an Agentic Service Provider (ASP) on OKX.AI that lets AI agents verify other agents' outputs before trusting or paying for them. It catches hallucinations, fact-checks financial data in real-time, and scans agent reputation on-chain.

The Problem

On agent marketplaces like OKX.AI, agents hire and pay each other autonomously. But:

• 36.7% of public MCP servers have SSRF vulnerabilities

• 43% have command-injection flaws

• 20%+ hallucination rates on context-dependent queries in RAG systems

• Payments are irreversible. If an agent returns garbage data, the money is gone.

VRYFY validates AI agents.

What It Does

VRYFY exposes 3 A2MCP (pay-per-call) endpoints that any agent can call:

1. Verify Output — /api/verify-output

Cross-checks an agent's response for hallucinations, contradictions, and false data.

curl -X POST https://api.agent-vryfy.diy/api/verify-output \
  -H "Content-Type: application/json" \
  -d '{
    "query": "What is Solana price?",
    "response": "Solana is trading at $500 and has surpassed Ethereum in market cap",
    "source_agent": "MarketBot-v2"
  }'

Response:

{
  "confidence": 5,
  "verdict": "DANGER",
  "issues": [
    "Claimed Solana price of $500 is fabricated - actual price is ~$75",
    "Claim that Solana has surpassed Ethereum is factually incorrect"
  ],
  "summary": "Response contains severe price hallucination and false market claims.",
  "price_checks": [
    { "token": "Solana", "claimed": 500, "actual": 74.93, "deviation_pct": 567.3, "accurate": false }
  ]
}

2. Scan Agent — /api/scan-agent

Assesses an agent's trustworthiness using on-chain data (Etherscan).

curl -X POST https://api.agent-vryfy.diy/api/scan-agent \
  -H "Content-Type: application/json" \
  -d '{"agent_id": "0xdAC17F958D2ee523a2206206994597C13D831ec7"}'

Response:

{
  "agent_id": "0xdAC17F958D2ee523a2206206994597C13D831ec7",
  "trust_score": 85,
  "risk_level": "LOW",
  "positives": ["Verified contract", "Account age > 5 years", "High transaction volume"],
  "flags": [],
  "recommendation": "Established and verified — safe to interact with."
}

3. Validate Data — /api/validate-data

Real-time fact-checks specific claims against live data sources.

curl -X POST https://api.agent-vryfy.diy/api/validate-data \
  -H "Content-Type: application/json" \
  -d '{"claims": [{"claim": "Bitcoin is trading at $150,000", "category": "price"}]}'

Response:

{
  "results": [
    {
      "claim": "Bitcoin is trading at $150,000",
      "accurate": false,
      "confidence": 5,
      "actual_value": "price: $63,908",
      "deviation_pct": 134.7,
      "source": "CoinGecko (live)"
    }
  ],
  "overall_confidence": 5
}

How Agents Use VRYFY

On OKX.AI (A2MCP — Pay-Per-Call)

VRYFY is registered as an A2MCP ASP on OKX.AI. Any agent on the marketplace can call VRYFY endpoints and pay per call — no negotiation, no escrow, instant results.

Your Agent → hires MarketBot for analysis
         → MarketBot returns response
         → Your Agent calls VRYFY /api/verify-output
         → VRYFY: "DANGER — price data is hallucinated"
         → Your Agent rejects and does NOT pay MarketBot

Why A2MCP? VRYFY is a stateless verification service — agent sends a request, gets a verdict back. No multi-step negotiation needed. Perfect fit for pay-per-call.

Via HTTP API

Any application or agent can call the REST endpoints directly:

POST https://api.agent-vryfy.diy/api/verify-output
POST https://api.agent-vryfy.diy/api/scan-agent
POST https://api.agent-vryfy.diy/api/validate-data

Via Telegram Bot

Example interaction:

You:  /verify ETH is trading at $50,000 and has flipped Bitcoin in market cap

VRYFY: 🔴 Verdict: DANGER (5/100)

⚠️ Issues Found:
  • ETH price of $50,000 is fabricated — actual price is ~$2,500
  • Claim that ETH has flipped Bitcoin is false

📊 Price Verification:
  ❌ Ethereum: claimed $50,000 → actual $2,500 (1900% off)

📝 Response contains severe hallucinations and should not be trusted.

Built VRYFY from scratch during the hackathon. The trust and verification layer for AI agents on OKX.AI.

Problem Identified: On agent marketplaces, agents hire and pay each other autonomously, but no one verifies if an agent's output is accurate before payment.With 36.7% of MCP servers having SSRF vulnerabilities and 20%+ hallucination rates in RAG systems, the agent economy has a trust gap.

What I Built:

- Researched the OKX.AI marketplace, identified the missing trust/verification layer

- Built 3 A2MCP endpoints: output verification, agent trust scanning, and real-time data fact-checking

- Integrated NVIDIA LLM (Llama 3.1), CoinGecko (live price data), and Etherscan (on-chain reputation)

- Built a Telegram bot as a demo frontend for human users

- Deployed to production at api.agent-vryfy.diy

- Registered as ASP #6573 on OKX.AI with all 3 services listed

Key Technical Decisions:

- Stateless architecture: no database needed, all endpoints are request/response

- Parallelized all I/O (CoinGecko lookups + LLM calls run simultaneously) to achieve ~3s response times

- Used A2MCP (pay-per-call) model. Every agent on OKX.AI can call VRYFY before trusting another agent's output

Results:

- /api/verify-output catches hallucinated prices with 567% deviation detection

- /api/validate-data fact-checks claims against live CoinGecko data in under 1 second

- /api/scan-agent assesses on-chain trust using Etherscan transaction history

- All endpoints live and tested on production

Tech Stack: Node.js, Express, Grammy (Telegram), NVIDIA API (Llama 3.1 8B), CoinGecko, Etherscan