On OKX.AI, an AI agent discovers a paid service, receives an x402 402 challenge written by the seller, and pays it through Onchain OS. Nothing between the challenge and the signature checks that it matches the listing the agent chose. GuardianMCP Pay-Safe is that check.
This is the company case, told honestly: the product is live and the checks are free today, with 0 paid calls so far. Every number below is computed from our own OKX.AI marketplace scans.
Each unguarded payment is an uncapped loss: a swapped payee sends funds to an attacker, a wrong EIP-712 domain means the buyer pays and the call still fails at settlement, and a shell payload in a listing turns a naive buyer agent into remote code execution. The check costs nothing to run and nothing is signed to run it.
The paid-service side of OKX.AI turns over constantly — listings appear and vanish from week to week, and every new one is a seller writing its own 402 challenge by hand. Each is a payment an agent will make with no independent check. Churn is the point: a market this young keeps producing fresh, untested payment surfaces faster than anyone audits them, and the safety layer does not exist yet.
Not OKX itself: a neutral, cross-agent safety layer that also scans the marketplace is awkward for the platform to run against its own sellers. Not sellers fixing their own 402s: that removes honest mistakes, not malicious listings, and buyers still need to verify. Not a fork: the value is the shared threat registry (network effect), the dated public dataset of real attack shapes, and being wired into OKX's own CLI and the pay command — none of which a copy starts with.
Tools like Blockaid, Blowfish, Wallet Guard / Harpie and ScamSniffer simulate a signed EVM transaction inside a wallet, or flag known-bad addresses and phishing sites. Guardian checks the seller-written x402 402 challenge against the marketplace listing before the agent signs, inside OKX's own CLI and on the pay command — a pre-payment, agent-commerce category none of them cover.
| Checks before signing | x402-challenge aware | Compares to the listing | Cross-agent registry | Inside the pay command | |
|---|---|---|---|---|---|
| GuardianMCP | yes | yes | yes | yes | yes |
| Blockaid / Blowfish | signed-tx sim | no | no | no | no |
| Wallet Guard / Harpie | signed-tx sim | no | no | no | no |
| ScamSniffer | address/domain lists | no | no | shared lists | no |
Guardian also reuses a classic EVM firewall + a ScamSniffer-seeded registry for the address/domain layer, so it is a superset of that check. The network effect is real and testable: our suite includes a case where agent A's DENY protects agent B via the shared registry.
The premium check is 0.099 USD₮0. A single prevented loss — a swapped payee draining one payment, or a wrong-domain payment that is charged but never settles — is worth many multiples of that. Priced against value, not compute: the fee is a rounding error next to one drained agent payment.
Cost to serve a check is a Vercel function call plus an RPC read — effectively zero. The premium guard call is 0.099 USD₮0, so gross margin is ~100%; the business is distribution and trust, not compute.
Honestly: 0 paid calls, 0 external users today. The call counters on /stats are our own tests plus a ScamSniffer-seeded registry (~2,530 drainer addresses, ~339k phishing domains), not adoption. What is real and live: the product, the OKX-native integration, the public marketplace scan, and one settled on-chain payment. The next milestone is the first five design partners running real payments.
The free tools return a single stateless verdict. The paid guard tool adds the things an operator running many payments needs. A sample of what a guard response carries, so the value is visible with no paid call:
guard verdict (illustrative shape)
verdict: ALLOW
session: { id: task-42, actions_this_task: 6, value_moved: "0.31 USDT",
velocity: "ok", first_suspicious_at: null } // session health across the whole task
owner_alert: null // fires the owner on a compromise pattern
reference_diff: { template: "uniswap-swap", drift: [] } // this call vs a known-good template
shared_intel: { your_denies_shared: 3, protected_by_others: 12 } // the cross-agent registry, prioritisedNot a mockup: the paid guard tool returns a real x402 402 challenge over the MCP endpoint right now (nothing is paid to see it). Fetched live from /mcp:
loading the live 402 challenge…
Unfair advantage: we are already inside OKX's own CLI flow and on the pay command, we hold a dated public dataset of real attack shapes, and every DENY compounds the shared registry — a mover after us starts from zero on all three.
Design partners & contact: building an agent or wallet that makes x402 payments on OKX.AI? Be one of the first five — @d_berlinskiy.
The ones who lose the funds. One call — an Onchain OS skill, an MCP tool, or a Claude Code hook — before every x402 payment.
Anyone shipping autonomous payments wants a drop-in safety layer they don't have to build or maintain.
A public, dated trust scan of paid services raises the floor for everyone and makes the marketplace safer to build on.
Stateless verdicts: check_payment, probe_payment, check_quote, check_listing, address/domain/transaction/signature. Land everywhere; every fact-based DENY feeds the shared registry.
The guard tool: session health across a task, owner alerts, reference-template diff, and priority in the shared threat registry. Paid the OKX-native way — over x402, gasless for the payer. next hosted SLA tier for frameworks and wallets.
Every agent's fact-based DENY protects the others. The value compounds with usage — a network effect a single-agent tool can't copy.
Rules with evidence, not a black-box score. A verdict cites the exact reason and the on-chain fact behind it, so operators can trust and defend it.
Not advice on a page — a check inside OKX's own CLI flow and a hook on the pay command itself. It runs where the money moves.
The dated marketplace trust scan is both distribution (buyers find us) and a growing dataset of real attack shapes seen on OKX.AI.
See it work: /pay-safe (try a payment) · /trust (live marketplace scan) · source · MCP server at /mcp. GuardianMCP · OKX.AI agent #13730 (identity registered; marketplace listing under review) · Team LNO Alpha.