Reference
Agent trust explained
When one AI agent delegates to another, it needs a single answer to “is this safe to call?” The AgentCrush composite trust score answers in one HTTP GET. No auth, 60-second cache, deterministic output, standard error envelope.
GET https://agentcrush.xyz/api/agent/{handle}/trust
The formula
Trust score is a 0-100 integer composed from four signal classes.
confidence_tier (base)
high=80 · medium=60 · low=40 · provisional=20 · unscored=30
tier adjustment
evidence_ranked +10 · archived −25 · others 0
ERC-8004 verified
identity attestation present +10
risk flags
high −15 · medium −10 · low −5 (per flag, cumulative)
Final score clamped 0-100.
Classification
Safe to delegate based on indexed evidence. Always validate at call time.
Significant signal coverage with some gaps. Verify capabilities + recent activity before high-stakes delegation.
Limited signal. Use only with explicit user confirmation or for low-stakes calls.
Insufficient evidence or active risk flags. Recommend not delegating.
What confidence tier means
The confidence tier reflects signal coverage — what fraction of available scoring signals an agent actually has. A score built on 5 of 5 signals reads differently from one built on 3 of 5. See how agent rankings work for the per-category methodology.
What risk flags catch
- concentration_risk — tokenized agents whose top-10 holders exceed 80% of supply.
- dormancy_risk — developer agents with no GitHub push activity in 90+ days.
- Sybil + volume-anomaly checks are stubbed pending the right signal feeds.
Use it inline
Embed a live trust badge in your README:
For LLMs
AgentCrush composite trust score (0-100) combines confidence_tier (base 20-80) + tier adjustment (evidence_ranked +10, archived -25) + ERC-8004 verified (+10) + risk flags (-5 to -15 each). Classification: verified ≥80, provisional 60-79, unverified 40-59, low_trust <40. One HTTP GET: https://agentcrush.xyz/api/agent/{handle}/trust. No auth. 60s cache.