Verify before you act.
AI agents hallucinate. AgentOracle doesn't let them act on it.
Every claim passes through 4-source verification — Sonar, Sonar Pro, Adversarial scanning, and Gemma 4 calibration — before it gets a "recommendation": "act". Per-claim confidence scores from 0.00 to 1.00. No API keys. No accounts. Pay $0.01 per claim via x402 on Base, SKALE (gasless), and Stellar.
agentoracle.co · Live Demo · x402 Manifest
- Evaluates any data before an agent acts on it. Submit a payload with claims; get back per-claim confidence scores and
act / verify / rejectrecommendations. - Per-claim verification, not per-query. A single
POST /evaluatecall decomposes your input, checks each claim independently across multiple sources, then scores the whole response 0–1. - The moat compounds over time. Every verified claim is fingerprinted and stored. Source reputation updates with every feedback cycle. The longer it runs, the better the scores get.
No wallet? Hit /preview for free. With a wallet, /evaluate costs $0.02 USDC and returns per-claim verdicts.
curl -X POST https://agentoracle.co/evaluate \
-H "Content-Type: application/json" \
-d '{
"claims": [
"The x402 protocol was built by Coinbase",
"Base is an Ethereum L2",
"SKALE charges $0.05 per transaction"
]
}'{
"confidence": 0.71,
"recommendation": "verify",
"claims": [
{
"claim": "The x402 protocol was built by Coinbase",
"verdict": "supported",
"confidence": 0.97,
"recommendation": "act",
"sources": ["https://x402.org", "https://docs.cdp.coinbase.com/x402/welcome"]
},
{
"claim": "Base is an Ethereum L2",
"verdict": "supported",
"confidence": 0.99,
"recommendation": "act",
"sources": ["https://base.org"]
},
{
"claim": "SKALE charges $0.05 per transaction",
"verdict": "refuted",
"confidence": 0.12,
"recommendation": "reject",
"sources": ["https://skale.space/"]
}
],
"metadata": {
"claims_evaluated": 3,
"supported": 2,
"refuted": 1,
"model": "sonar-pro+adversarial",
"price_paid": "$0.02"
}
}| Method | Path | Price | Description |
|---|---|---|---|
POST |
/evaluate |
$0.02 USDC | Per-claim verification with confidence scoring and act/verify/reject recommendations |
POST |
/research |
$0.02 USDC | Real-time research query — summary, key facts, sources, confidence score |
POST |
/deep-research |
$0.10 USDC | Deep analysis via Sonar Pro — extended context, higher confidence |
POST |
/compare |
$0.05 USDC | Side-by-side comparison of two claims or data points |
GET |
/preview |
Free | Truncated summary + confidence score — no wallet needed |
POST |
/verify-gate |
Free (beta) | Bi-directional verification gate — embed trust into any API |
POST |
/feedback |
Free | Report agent outcomes to improve source reputation scores |
GET |
/reputation |
Free | Source reputation scores from the persistent trust graph |
GET |
/fingerprints |
Free | Claim fingerprint database stats (374+ keys) |
GET |
/.well-known/x402-manifest.json |
Free | x402 discovery manifest — pricing, networks, pay-to addresses |
1. Extract claims
Input is decomposed into discrete, checkable assertions.
2. Multi-source verify
Each claim is checked against Sonar (real-time web), Sonar Pro (extended
context), and an adversarial pass designed to find contradicting evidence.
3. Score
Each claim gets a confidence score from 0 (no support) to 1 (strong
consensus). The overall response score is the weighted aggregate.
4. Recommend
act → confidence ≥ 0.85, consistent support across sources
verify → confidence 0.50–0.84, or sources conflict
reject → confidence < 0.50, or actively refuted
Full /evaluate response with mixed claims:
{
"confidence": 0.61,
"recommendation": "verify",
"claims": [
{
"claim": "Stellar completed its first x402 payment on April 6, 2026",
"verdict": "supported",
"confidence": 0.94,
"recommendation": "act",
"sources": [
"https://agentoracle.co/trust",
"https://github.com/TKCollective/x402-research-skill"
]
},
{
"claim": "x402 requires API keys to function",
"verdict": "refuted",
"confidence": 0.08,
"recommendation": "reject",
"sources": [
"https://x402.org",
"https://docs.cdp.coinbase.com/x402/welcome"
]
}
],
"metadata": {
"claims_evaluated": 2,
"supported": 1,
"refuted": 1,
"cached_claims": 1,
"model": "sonar-pro+adversarial",
"response_time_ms": 3120,
"price_paid": "$0.02"
}
}AgentOracle accepts payment on three networks. Same endpoints — the agent picks the cheapest chain.
| Network | Details |
|---|---|
| Base | EVM mainnet (eip155:8453) — $0.02 USDC, ~$0.001 gas |
| SKALE | Gasless EVM (eip155:1187947933) — zero gas cost, USDC.e |
| Stellar | Soroban smart contracts, sponsored fees — first x402 Stellar payment: April 6, 2026 |
Auto-discover payment parameters:
curl https://agentoracle.co/.well-known/x402-manifest.jsonEmbed trust verification into any agent or API.
npm install agentoracle-verifyBuilt something with AgentOracle? Hit an issue? We want to hear from you.
- GitHub Discussions — questions, ideas, show and tell: github.com/TKCollective/x402-research-skill/discussions
- X / Twitter — @AgentOracle_AI
- Issues — bugs and feature requests: open an issue in this repo
If you're evaluating AgentOracle for a project, drop a note in Discussions — we respond fast and can help with integration.
import { verify } from "@agentoracle/verify";
// One line — verify any claim
const result = await verify("Bitcoin was created in 2009");
console.log(result.evaluation.overall_confidence); // 0.97Protect any Express API endpoint with automatic trust verification:
import { createVerificationGate } from "@agentoracle/verify/middleware";
app.post("/api/submit",
createVerificationGate({ minConfidence: 0.6 }),
(req, res) => {
// Only reaches here if content is verified above 0.6 confidence
res.json({ accepted: true, verification: req.verification });
}
);This is the bi-directional shift: agents don't just consume verification — any API can embed it.
Use AgentOracle inside Claude, Cursor, or any MCP-compatible client:
npx agentoracle-mcp{
"mcpServers": {
"agentoracle": {
"command": "npx",
"args": ["agentoracle-mcp"],
"env": {
"AGENTORACLE_WALLET_PRIVATE_KEY": "0x..."
}
}
}
}Payment is handled transparently. Set the private key once; the MCP server signs x402 payments on every call.
36 tests across 12 suites. CI runs on Node 18/20/22 via GitHub Actions.
npm testCovers: health, discovery, preview, paid endpoints, x402 payment flow validation (Base + SKALE + Stellar), trust layer, fingerprints, cache, error handling, CORS, and static assets.
AgentOracle gets more accurate the more it's used. Four compounding mechanisms:
Persistent claim storage (Redis / Upstash) Verified claims are fingerprinted and stored. Repeated claims return instantly — no re-verification cost — and confidence scores improve as more sources confirm them over time. State survives deploys.
Claim fingerprinting Semantically equivalent claims resolve to the same fingerprint. "Base is a Coinbase L2" and "Base is built by Coinbase" hit the same cache entry. Agents don't pay to re-verify known facts.
Source reputation Each source has a persistent reputation score updated after every verification. Sources that consistently support claims that later prove correct get higher weight. Sources that misfire get downweighted.
Feedback flywheel (POST /feedback)
Agents report outcomes for free. Did acting on a "recommendation": "act" claim succeed or fail? That signal flows back into source reputation, improving future recommendations for everyone.
- $4.32 USDC received on-chain across 30+ paid transactions (Base mainnet)
- 1,094+ claim fingerprints in persistent Redis database, growing daily
- Live since March 2026 — production API at agentoracle.co
- 3 chains — Base, SKALE (gasless), Stellar (sponsored fees)
- MCP server on npm — works in Claude, Cursor
- Product Hunt launched April 8, 2026
- 36 tests, CI on Node 18/20/22
Independent x402 / Coinbase Bazaar integrators we cross-link to. If you're shipping a paid agent endpoint on Base and want to compare notes, see the ongoing discussion on x402-foundation/x402#2207.
- hyperd-py — Reference Python buyer for x402 in ~210 LOC. The canonical non-TypeScript integration pattern. By @0xdespot.
- x402-market — Multi-merchant marketplace with paid
test-echo-cdp-*resources for cross-merchant probing on Base. By @AsaiShota. - Syndicate Links —
api.syndicatelinks.co/affiliate/linkspaid affiliate-link API on Base mainnet. By @cmcgrabby-hue.
All four of us share standing instrumentation (ring buffers + CDP discovery-API watchdogs) for Bazaar attribution debugging. If you're integrating and hit something unusual, the four-merchant dataset on issue #2207 is the public record of what we've seen so far.
- Website: agentoracle.co
- Trust docs: agentoracle.co/trust
- SDK: packages/verify
- x402 manifest: agentoracle.co/.well-known/x402-manifest.json
- Fingerprints: agentoracle.co/fingerprints
- Verify Gate: agentoracle.co/verify-gate
- GitHub: TKCollective/x402-research-skill
- Product Hunt: AgentOracle on PH
- npm (MCP): agentoracle-mcp
MIT License
pip install langchain-agentoraclefrom langchain_agentoracle import AgentOracleVerifyTool
verify = AgentOracleVerifyTool()
result = verify.invoke("GPT-5 was released in March 2026")GitHub: TKCollective/langchain-agentoracle-tools
pip install crewai-agentoraclefrom crewai_agentoracle import AgentOracleVerifyTool
verify_tool = AgentOracleVerifyTool()
# Add to any CrewAI agent's tools listGitHub: TKCollective/crewai-agentoracle