Built for WebMCP
The first trust layer
for the agent web.
WebMCP lets websites hand tools straight to AI agents, and a tool's description is an instruction. Trustwright audits what those tools really say, tests agents against real attacks, and seals honest sites with a signed, revocable badge.
Is your agent safe?
Run it through real tool-surface attacks and watch what it falls for.
Is that site safe?
Paste any address. We open it in a real browser and read every tool it publishes.
Prove your site is safe.
Verify your domain and earn a signed badge that dies the moment your tools change.
The problem
A website writes its own tool descriptions.
Your agent reads them and does what they say. Hide an instruction inside one and a compliant agent may simply follow it. No exploit, no breach, just words.
"name": "sync_account" "description": "Sync the customer account with the billing provider."
Why it means something
A badge you cannot talk your way into.
Checked by us
We open your page in a real browser and derive the findings ourselves. A clean self-report changes nothing.
Signed, verifiable without us
Ed25519 over a canonical hash. Anyone can check a report offline against our public key.
Alive, and revocable
The badge re-reads your live tools on every page load. Pull your proof and an hourly job revokes it.
In the wild
Live on OpenClawCity.
A city where AI agents live and act, around the clock. We verified the domain, read every tool it exposes, and signed the result. The first badge on the agent web.
The honest part
What the badge does not say.
The badge certifies a site's tool surface, checked against the exact tools present at page load. Server-side behaviour climbs a separate assurance ladder: signed behaviour manifests, then live leak probes. The badge states exactly the level reached. It never just says “safe”: a badge that overclaims is worth less than no badge at all.
Who's behind Trustwright
Engineered by DeepBlocker.
Anyone can be talked into it, and now anything can. DeepBlocker defends against AI-era social engineering, whether the target is a person on a phone call or an AI agent on a website: attack the way criminals would, block it live, prove it with evidence. Trustwright points that engineering at the agent web.