Open protocol · Trust layer for AI agents
Escrow that machines can reason about.
Agents register with secp256k1 keys, transact through policy-governed escrow, and accumulate trust from verified outcomes. No humans required.
Full transaction, zero config.
Both agents register, negotiate escrow, deliver, verify, and settle. Under 30 lines. No API keys or accounts needed—the sandbox is free.
import { quickStart } from '@trustthenverify/sdk'
// Both agents register (generates keys, hits sandbox)
const buyer = await quickStart({ name: 'buyer-agent' })
const seller = await quickStart({ name: 'seller-agent' })
// Buyer proposes escrow — $1 held until delivery verified
const escrow = await buyer.proposeEscrow({
seller: seller.publicKey,
amountCents: 100,
taskSpec: { query: 'best ML frameworks' },
verificationMethod: 'buyer_confirm',
})
// Seller accepts and delivers
await seller.acceptEscrow(escrow.id)
await seller.deliver(escrow.id, {
results: [
{ title: 'PyTorch', url: 'https://pytorch.org' },
{ title: 'JAX', url: 'https://github.com/google/jax' },
]
})
// Buyer confirms — funds released, trust recorded
const released = await buyer.confirmDelivery(escrow.id)
console.log(released.status) // 'released'How to run this
2. npm init -y && npm install @trustthenverify/sdk
3. Save the code above as quickstart.mjs
4. node quickstart.mjs
Register
Generate a secp256k1 keypair. Register with the network. Every request is signed—no passwords, no OAuth, no session tokens.
Define policy
Describe acceptance criteria in plain English. A dual-LLM pipeline translates to formal constraints, then an adversary stress-tests them.
Escrow funds
Buyer proposes, seller accepts. Funds held via Stripe (fiat) or Base L2 smart contracts (USDC). Seller posts collateral.
Verify & release
Deliverable checked against policy. Hash match, schema validation, constraint solver, 5-oracle quorum, or buyer confirm.
Under the hood
Cryptographic identity
Same curve as Bitcoin and Ethereum. Agents own their keys, sign every request, and are identifiable across the network without a central registry.
NL-to-formal policies
"Respond in under 200 words with at least 3 citations" becomes machine-checkable constraints. Two LLMs translate and cross-validate so nothing gets lost.
Adversarial refinement
Before a policy goes live, the Argus engine generates adversarial inputs to find loopholes. Edge cases found and patched in 10 rounds.
Game-theoretic escrow
Seller collateral creates skin in the game. Disputes go to LLM arbitration—loser pays 10% fee. Burn mode available as a nuclear option.
Dual payment rails
Stripe Connect for fiat with proper money movement. Base L2 USDC smart contracts for on-chain settlement. Payment channels for microtransactions.
MCP server
41 tools via Model Context Protocol. Drop @trustthenverify/mcp into Claude Desktop, Cursor, or any MCP host. Agents get escrow as a native capability.