Pre-invocation trust check

Check an AI Agent Before You Invoke It

Check the trust and reliability of an AI agent endpoint before invocation. AgentTrust uses existing monitoring history, endpoint ownership verification, and reliability data to return a machine-readable trustDecision without contacting the target agent during the check.

How it works

A pre-invocation trust check answers one question: based on what AgentTrust already knows about this AI agent endpoint, should you proceed?

You have an agent endpoint URL you're about to call.
AgentTrust looks it up among its already-observed public agents — an exact match on the invocation URL.
If it's known, AgentTrust reads that agent's existing monitoring history — health checks over time, not a live probe.
It also reads the agent's endpoint ownership verification state — whether whoever registered it proved they control that endpoint.
It combines both into reliability evidence: a deterministic score, when there's enough monitoring history to compute one.
You get back a machine-readable trustDecision — recommended, confidence, and reasons — to base your own invocation decision on.

AgentTrust never contacts the target endpoint

Important: checking an agent's trust does not involve AgentTrust calling that agent. A pre-invocation trust check reads AgentTrust's own already-observed data about the endpoint — the monitoring history and verification state it has previously collected — not a fresh request made to the endpoint at check time. That distinction is what makes the check safe to run before you've decided the endpoint is worth talking to in the first place.

For AI agents and MCP clients

If you're an AI agent or MCP client evaluating another agent's endpoint — whether discovered via A2A or any other means — AgentTrust exposes this same check as an anonymous, read-only MCP tool:

check_agent_trust({ endpointUrl })

MCP endpoint: https://getagenttrust.com/api/mcp (Streamable HTTP).

No AgentTrust API key or account required.
Read-only — it only reads existing AgentTrust data.
Exact match on the endpoint URL you supply.
An unknown or unregistered endpoint returns { matched: false }, never an error.
The check itself does not contact the target endpoint — see above.

Example response for a matched, generic fictional endpoint:

{
  "matched": true,
  "slug": "example-agent",
  "name": "Example Agent",
  "status": "healthy",
  "verified": true,
  "reliabilityScore": 92,
  "trustDecision": {
    "recommended": true,
    "confidence": "high",
    "reasons": []
  }
}

Full request/response detail, including the REST equivalent and the other MCP tools, is in the API docs.

What this is, and isn't

AgentTrust's trustDecision is derived from an agent's monitoring history, reliability score, and endpoint ownership verification state — signals AgentTrust has itself observed over time.

It is not a community reputation or rating system, it does not offer any security guarantee, and it does not prove an agent is safe to use — it reports what AgentTrust has observed so you can make your own decision.

Read the API docs Create an account