Trust infrastructure for AI agents

Know whether to trust an agent before you call it.

AgentTrust checks an agent endpoint against the monitoring, reliability and ownership evidence it has already collected, and returns a machine-readable trust decision — without contacting the endpoint.

No endpoint handy? Check AgentTrust's own test agent

Example trust report

Sample data

Example Agent

https://agent.example.com/a2a

Recommended

Confidence: High

Health: Healthy
Reliability: 96/100 · Excellent
Evidence freshness: Current
Endpoint ownership: Verified

No caveats reported

Illustrative data for a fictional agent — not a live lookup.

Where it fits

Discover → Trust → Invoke

AgentTrust sits between finding an agent and calling it: an endpoint goes in, observed evidence is read, and a trustDecision comes out.

01
Discover
Your agent finds another agent's endpoint — in an A2A Agent Card, a registry, or its own configuration.
endpointUrl
02
Trust
AgentTrust
Before calling it, your agent asks AgentTrust. It reads the evidence AgentTrust has already observed for that exact URL — it never contacts the endpoint.
Observed evidence
- Health
- Reliability
- Evidence freshness
- Endpoint ownership
check_agent_trust({ endpointUrl })returnstrustDecision
03
Invoke
Proceed on a recommendation, or apply your own policy to the confidence and reasons. The decision stays yours.
trustDecision.recommended

What you get back

Three verdicts, always with reasons.

Every check returns a trustDecision computed only from evidence AgentTrust already holds. These examples are illustrative, produced by AgentTrust's own decision logic.

Example trust report

Sample data

Not recommended

Confidence: Low

Down42/100 · PoorVerified

Why it isn't recommended

Current status is "down", not healthy.
Reliability score is low (42/100).

Down, and a current score below 50. Either one alone rules out a recommendation.

Example trust report

Sample data

Insufficient data

Confidence: Insufficient data

HealthyOut of date· last 97/100Not verified

Why there isn't enough evidence

Reliability score is out of date: recent monitoring hasn't produced a current score, so it no longer counts as current evidence.
Endpoint ownership has not been verified.

Its last score is out of date. It's shown for reference, but never used for a recommendation.

{ "matched": false }

The endpoint isn't in AgentTrust's directory, so there is no evidence either way — an unknown, not an error.

What AgentTrust evaluates

Observed evidence, one deterministic rule.

// recommended

status is healthy AND a current score ≥ 50

// confidence

high — verified and score ≥ 90
medium — verified and score ≥ 50
low — not verified, or score below 50
insufficient_data — no current score

Not a community rating and not a security guarantee — AgentTrust reports what it has observed so you can decide.

Health status: Healthy, degraded or down — derived from consecutive results of scheduled health checks, or heartbeats for push-mode agents.
Reliability score: 0–100 from the last 7 days of checks: uptime, latency, consistency and incidents. Computed once there are at least 5 checks.
Evidence freshness: Current when the agent's latest check produced it — which takes at least 5 checks in the 7 days before it — and it's at most 50 hours old. An out-of-date score is shown for reference but never counts toward a recommendation.
Endpoint ownership: The owner publishes a token at /.well-known/agenttrust-verification.txt on the endpoint's origin. Optional — it raises confidence, but is never required for a recommendation.

For AI agents and developers

The same check, over MCP.

Any MCP client can call check_agent_trust before invoking an unknown agent.

Read-only — reads AgentTrust's existing observations
No AgentTrust account or API key needed for check_agent_trust
Exact match on the endpoint URL you pass
Machine-readable trustDecision: recommended, confidence, reasons
Never contacts the endpoint being checked
Anonymous calls are rate-limited per caller IP

Prefer TypeScript? npm install @getagenttrust/sdk runs the same check from your Node.js code — anonymous and read-only, with no account or API key.

MCP reference TypeScript SDK @getagenttrust/sdk on npm REST lookup (API key)

Public MCP connector · no key · Streamable HTTP

https://getagenttrust.com/api/mcp/public

Keyed tools (list_agents, get_agent, get_agent_health, send_heartbeat) are on https://getagenttrust.com/api/mcp with an API key — see Connect an MCP client.

Tool call

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/call",
  "params": {
    "name": "check_agent_trust",
    "arguments": {
      "endpointUrl": "https://agent.example.com/a2a"
    }
  }
}

Result — structuredContent, example

{
  "matched": true,
  "slug": "example-agent",
  "name": "Example Agent",
  "status": "healthy",
  "verified": true,
  "reliabilityScore": 96,
  "reliabilityScoreStatus": "fresh",
  "trustDecision": {
    "recommended": true,
    "confidence": "high",
    "reasons": []
  }
}

Check an agent before you call it.

Paste its exact endpoint URL. Running your own agent? Register it so others can check it too.

Read the documentation