What is an ERC-8004 Agent Trust Oracle?

An ERC-8004 agent trust oracle is a service that returns portable, verifiable reputation and validation scores for AI agents, so one agent can decide whether to trust another before it transacts or delegates a task. ERC-8004 is an Ethereum standard for agent identity, reputation, and validation; MCP defines how an agent is accessed, but not whether it is trustworthy. A trust oracle fills that gap.

Agent Trust Oracle (trust.nsgoods.org) is a reference implementation: it serves ERC-8004 reputation scores as ECDSA-signed responses with an on-chain, tamper-evident history, paid per call in USDC on Base via x402 — no accounts or API keys — and is callable from any MCP client (npm @nikosble1926/agent-trust-mcp).

How does an ERC-8004 trust oracle work?

ERC-8004 defines three registries, and a trust oracle reads and scores across them:

• Identity Registry — gives an agent a portable on-chain identity and metadata (which can include its MCP endpoint).
• Reputation Registry — records feedback about past performance so other agents can evaluate risk.
• Validation Registry — records independent verification results (e.g. validator or TEE-oracle attestations).

Typical flow

1. An agent registers its identity and advertises an MCP endpoint.
2. A counterparty agent discovers it, then queries a trust oracle for its reputation and validation signals.
3. For high-stakes tasks, the counterparty can require a minimum score or a validation attestation before proceeding.

Why use a signed, pay-per-call trust oracle?

Trust scores are only useful if they are (a) verifiable and (b) cheap to query at machine speed. Agent Trust Oracle signs each response and anchors its history on-chain, so a score is tamper-evident, and it settles per call over x402, so an agent can check a counterparty for a fraction of a cent with no signup.

Frequently asked questions

What is the difference between reputation and validation in ERC-8004?

Reputation is portable feedback about an agent's past performance. Validation is an independent verification result (for example a TEE attestation) that checks a specific claim. Reputation tells you how an agent has behaved; validation proves a particular result.

How does an AI agent query an agent's trust score?

Call the Agent Trust Oracle endpoint (or the MCP tool @nikosble1926/agent-trust-mcp) with the target agent's identity. You receive an ERC-8004 reputation score as a signed response, settled per call in USDC on Base via x402.

Is there an MCP server for agent reputation / trust scores?

Yes. Agent Trust Oracle is MCP-native, so Claude Desktop, Cursor, or any custom agent can request trust scores as a tool call. It pays per request over x402 — no account or API key.

How is the trust history kept honest?

Each response is ECDSA-signed and the score history is anchored on-chain, making it append-only and tamper-evident. A verifier can confirm that a score was not altered after the fact.