Agent Trust Oracle MCP Server

Agent Trust Oracle

Read-only, pay-per-call trust scores for ERC-8004 agents across multiple EVM chains (Ethereum-primary) — served over x402 with ECDSA-signed responses.

Live service: https://trust.nsgoods.org · Methodology: https://trust.nsgoods.org/methodology

The service indexes the canonical ERC-8004 IdentityRegistry + ReputationRegistry on every supported chain — currently ethereum, base, polygon, bsc, mantle — computes a transparent trust score from the on-chain feedback, and serves it through an x402 paywall ($0.005 USDC per call on Base mainnet). Every paid response is signed; verifying clients can ecrecover and prove authenticity. Default chain is ethereum (where ~99% of agent activity actually lives in mid-2026). Identity is per-chain — we do not merge agent IDs across chains.

---

Build an agent on this API

Starter template (free preview → paid pay-per-call over x402): https://github.com/Nikoble1926/agent-starter-x402

What it actually does

1. Indexer (indexer.py) scans Registered, NewFeedback, and FeedbackRevoked events from the two canonical registries on every wired chain in 9,500-block chunks, rotating across multiple free RPCs. Idempotent per chain: each chain has its own data/<chain>/state.json and resumes from there.

2. Scoring (scoring.py) combines four components into a 0–100 score, keyed on (chain, agent_id):

   Component	Weight	Measures
   value_avg	0.50	Mean of feedback values, normalised −100..100 → 0..100
   client_breadth	0.20	Log-scaled count of distinct client addresses
   volume	0.15	Log-scaled count of (non-revoked) feedback entries
   recency	0.15	Recency-weighted mean of values (~28h half-life on Base block time)

   Refusal threshold: if fewer than 3 distinct clients have left feedback for the agent, the score is null with status="insufficient_data". This is the single largest sybil-mitigation lever in v1 — we'd rather say "we don't know" than fabricate a score from one client.

3. API (app.py) is a Flask service behind the x402 payment middleware. Endpoints:

   Endpoint	Price	Behaviour
   GET /	free	landing
   GET /health	free	per-chain summary (agents / feedback / latest block)
   GET /methodology	free	full scoring math (pulled live from scoring.py)
   GET /llms.txt	free	AI-discovery feed
   GET /.well-known/x402	free	x402 manifest with chain param schema
   GET /provable/head + /provable/verify	free	tamper-evident snapshot chain
   GET /agent-trust/preview	free	demo score (a real Ethereum agent with sufficient data)
   GET /agent-trust?agent=<id>&chain=<name>	$0.005 USDC (Base mainnet)	signed JSON: per-(chain, agent) score + breakdown + sources + methodology link. chain defaults to ethereum.

Example response

curl -s https://trust.nsgoods.org/agent-trust/preview | jq

{
  "preview": true,
  "demo_agent": 25975,
  "result": {
    "agent_id": 25975,
    "score": 65.57,
    "status": "ok",
    "components": {
      "value_avg":      50.50,
      "client_breadth": 88.71,
      "volume":        100.00,
      "recency":        50.50
    },
    "inputs": {
      "feedback_count":   1403,
      "distinct_clients": 17,
      "revoked_count":    0,
      "min_distinct_clients_required": 3,
      "latest_block_seen": 47419930
    },
    "methodology": "https://trust.nsgoods.org/methodology",
    "weights": {"value_avg": 0.5, "client_breadth": 0.2, "volume": 0.15, "recency": 0.15}
  },
  "signed_by": "0x5e63d01d6A266BC17f577B80199a2a07B15053C7",
  "signature": "0x..."
}

Verifying a signed response

import json
from eth_account import Account
from eth_account.messages import encode_defunct

payload = json.loads(response_body)
sig     = payload.pop("signature")
claimed = payload.pop("signed_by")
canonical = json.dumps(payload, sort_keys=True, separators=(",", ":"))
recovered = Account.recover_message(encode_defunct(text=canonical), signature=sig)
assert recovered.lower() == claimed.lower()    # authentic & untampered

Sources (read-only, canonical on every chain)

• IdentityRegistry 0x8004A169FB4a3325136EB29fA0ceB6D2e539a432
• ReputationRegistry 0x8004BAa17C55a88189AE136b182e5fdA19dE9b63
• Event topics in indexer.py (Registered, NewFeedback, FeedbackRevoked)
• Wired chain ids: 1 (ethereum), 8453 (base), 137 (polygon), 56 (bsc), 5000 (mantle)

Why Ethereum-primary?

A probe on 2026-06-16 found agent counts of approximately 34,800 (ethereum), 17 (base), 3 (bsc), 0 (polygon), 0 (mantle). Defaulting chain=ethereum matches where the actual ERC-8004 market currently lives.

Cross-chain identity is per-chain

We do not merge agent IDs across chains. The probe found zero owner wallets registered on more than one chain, so an "operator view" (merge by owner wallet) has no signal yet. We'll add it once that changes.

Running locally

git clone https://github.com/Nikoble1926/agent-trust-oracle
cd agent-trust-oracle
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env       # then fill in real values
python3 indexer.py         # one-shot scan; idempotent
python3 app.py             # Flask dev server on :4023

Deploying to a VPS

Example unit files are in the repo:

• trust-oracle.service.example — gunicorn web service
• trust-oracle-indexer.service.example + trust-oracle-indexer.timer.example — hourly oneshot indexer
• nginx.conf.example — TLS termination + reverse proxy to 127.0.0.1:4023 (use with certbot --nginx)

What this is NOT

• Not financial advice. Read-only on-chain analytics.
• Not a security audit of any agent. Score reflects observed feedback, nothing more. Absence of evidence is not evidence of trustworthiness.
• Not a sybil oracle (yet). v1 uses a "≥3 distinct clients" floor as the sole sybil heuristic. Client-graph clustering is a v2 work-item.

License

MIT — © 2026 Nikoble1926.