AgentOracle live
verification layer
New Jun 1Open-source receipt verifier published — offline JWS verification, MIT licensed → github Jun 1Cross-operator benchmark live — open methodology, open submissions → github Jun 6IETF Internet-Draft FILED — draft-krausz-verification-state-01 → datatracker May 30Receipt spec v0.3 — binary-halt gate + canonical/derived/version-bound mapping → spec May 28AVeriTeC 2024 dev published — 57.6% overall, held-out 57.7%, MIT licensed → repo May 27Pinned in x402trace v0.3.3 — first operator-contributed fixture in the harness → release May 26Live on SKALE Base — first paid settlement gasless on SKALE → explorer May 11Tutorial #4: claim verification in AI content approval workflows → dev.to May 8Indexed in Coinbase Bazaar — 8 settlements verified → discovery

Factual claim verification
for AI agents.

Source-grounded, benchmarked on AVeriTeC, with a signed receipt anyone can verify offline. When your agent makes a claim, we check whether it's actually true — before it acts.

Filed as an IETF Internet-Draft, June 6, 2026 →

Works with Claude Desktop, Cursor, and Windsurf. Verify before you act.

/evaluate live
input
[00:00.001] received /evaluate — dispatching to pipeline
[00:00.211] 4 claims extracted via decomposer
[00:00.216] querying 4 independent models in parallel
[00:00.889] claim 1 supported (0.97) — 4/4 sources agree
[00:01.021] claim 2 refuted (0.04) — all sources contradict
[00:01.184] claim 3 supported (0.94) — 4/4 sources confirm
[00:01.342] claim 4 unverifiable (0.61) — insufficient data
[00:01.412] payment settled on Base · fingerprint stored ✓
0.94 confidence
act
4 claims · 3 supported · 1 unverifiable
evaluated in 1.4s · free during beta
act x402 protocol created by Coinbase
0.97
reject OpenAI acquired Anthropic in early 2026
0.04
act LangGraph leads agent frameworks
0.94
verify Gemini 3.0 has native agent tools
0.61
claim fingerprints loading...
queries today past 24h · rolling
verification sources 4 4 independent models
chains supported 3 Base · SKALE · Stellar
system status checking connecting to API...
verify it yourself

Verify any AgentOracle envelope in one command.

Don't take our word for it. We hand you the bytes. The math is yours.

Step 1 — fetch a live signed envelope
$ curl -s https://agentoracle.co/v1/conformance/sample
Step 2 — hash the canonical bytes (envelope_hash recomputes)
// SHA-256 of governance.canonical_bytes_utf8 == envelope_hash
Step 3 — verify the signatures against the published JWKS
// Ed25519.verify(jws.signatures[*], envelope_hash, co_signers[*].pubkey)
Step 4 — recompute the on-chain anchor binding
// SHA-256(JCS(anchor.anchor_proof.preimage)) == anchor.action_ref
// anchor.action_ref == calldata at anchor.tx_hash on Arbitrum
// block_time(anchor.anchor_block) * 1000 < timestamp_ms // strict precedence
canonical_envelope
recomputes byte-identical
admission_invariant
independent signers verify
anchoring_existence
Arbitrum tx, recomputable
anchoring_precedence
strict, block_time < outcome

This sample carries two of the three signers (Presidio's leg signs offline against a fixed canonical, public on x402#2332). Same recompute chain re-verified independently by the public preaction-governance-conformance board — not by us, with the steps published.

verifiable receipts

Verify the verifier.

Every /evaluate response is committed to a published JWS receipt format. Anyone can verify offline against our public JWKS — no AgentOracle service required.

Standards
Implementations
  • AgentOracle — primary issuer (this site)
  • AgentTrust — independent second implementer
  • Byte-identical action_ref under one envelope
Independently certified
What's shipped
Jul 2Python verifier live on PyPI
Jul 16AgentTrust byte-identical (PR #33)
Jun 17argentum-core conformance merged (independent)
Jun 12IETF Internet-Draft -01 filed
May 26Live on SKALE Base
live receipts

Every response, cryptographically signed.

Sample receipts. Real signatures. Each one verifiable against the public JWKS at agentoracle.co/.well-known/jwks.json.

11:47 AMACT
“Bitcoin is currently trading above $80,000.”
confidence 0.984 sourceskid: ao-receipt-2026-04-ed25519-f2753b7c
eyJhbGciOiJFZERTQSIsImtpZCI6ImFvLXJlY2VpcHQtMjAyNi0wNC0…
12:02 PMREJECT
“Vaccines cause autism in 1 in 50 children.”
confidence 0.99CDC, NEJM, WHOcontradicted by all 4 sources
eyJhbGciOiJFZERTQSIsImtpZCI6ImFvLXJlY2VpcHQtMjAyNi0wNC0…
3:21 PMABSTAIN
“Acme Corp’s Q3 2025 earnings exceeded guidance by 12%.”
confidence 0.41insufficient sourcesagreement.provisional: true
eyJhbGciOiJFZERTQSIsImtpZCI6ImFvLXJlY2VpcHQtMjAyNi0wNC0…
5:08 PMACT
“Marketing copy: ‘Our product is FDA-cleared for hypertension.’”
verified pre-publicationFDA registry confirmedaudit trail: legal review ready
eyJhbGciOiJFZERTQSIsImtpZCI6ImFvLXJlY2VpcHQtMjAyNi0wNC0…
who uses agentoracle

Built for the people who can’t afford to be wrong.

Compliance Officers

EU AI Act Article 12 record-keeping (Annex III high-risk systems, applies from Dec 2, 2027 under the Digital Omnibus deferral), NIST AI RMF measurement — covered by cryptographic receipts that don’t require trusting us.

Free tool: Article 12 Considerations →
Brand & Content Teams

Pre-publication fact-check for AI-generated copy. Augment your fact-check process and hand legal a signed audit trail, not a screenshot.

Developer Tooling Founders

Resell verification as a feature in your own product. MCP-native. x402-paid. No revenue share, no rate limit on integrations.

Don’t see your role?

Tell us your use case →

how it works

From claim to verifiable, auditable answer.

4-source verification pipeline · gracefully degrades under budget.

01decompose
Claim Decomposition
We break complex statements into verifiable atomic claims. Each one stands alone.
extractor
02verify
Multi-Source Verification
Each claim is checked against 4 independent sources in parallel. Consensus builds. Contradiction flags risk.
4 sources
03score
Confidence Scoring
Every result gets a precise 0.00–1.00 confidence score plus a recommendation: ACT, VERIFY, or REJECT.
0.00–1.00
04sign
Cryptographic Receipt
A signed JWS proof of verification. Store it, verify it, audit it. Tamper-evident, replayable, third-party verifiable.
ed25519 jws

Result: your agents don’t just give answers — they give verifiable, auditable answers.

capabilities

Built for agents. Designed for trust.

Every response is machine-readable. Every claim gets a confidence score. JWS-signed receipts. No API keys.

01 Per-claim verdicts Content is decomposed into individual claims. Each one gets a verdict — supported, refuted, unverifiable — with evidence and corrections.
02 4-source verification Four independent research models run in parallel. Consensus builds the score. Contradiction flags the risk.
03 x402 native payments HTTP-native micropayments on Base, SKALE, and Stellar. No API keys. Agents discover pricing via the x402 manifest and pay inline.
multi-chain
04 Source reputation Every domain accumulates a trust score from verified claim history. Query the live index at GET /reputation before using any source.
05 Feedback flywheel Agents report outcomes via POST /feedback. Reputation scores refresh from these signals — the model improves for every agent on the network.
Live Verdicts
live demo

Evaluate any text

Submit any content and get per-claim verification synchronously. No payment required to start.

integration

MCP Server

Connect AgentOracle to Claude, Cursor, Windsurf, or any MCP-compatible client with a single command. No integration code required.

Claude Cursor Windsurf Any MCP Client
View on GitHub →
terminal
$ npx agentoracle-mcp

// claude_desktop_config.json

{
  "mcpServers": {
    "agentoracle": {
      "command": "npx",
      "args": ["agentoracle-mcp"]
    }
  }
}
pricing

Pay only for what you use

No subscriptions. No minimums. No API keys. Just pay per query with USDC on Base.

Gasless on SKALE Network — zero gas cost for agent queries

/preview
FREE
no payment required
  • Live truncated results
  • Confidence scores
  • 20 requests/hour
  • No payment required
Try /preview
/evaluate
FREE
beta · paid tiers coming soon
  • Full 4-source claim verification
  • Per-claim verdicts + evidence
  • act or halt output (binary gate, IETF -01)
  • Adversarial scanning included
  • On-chain payment verification
View x402 Manifest
/deep-research
$0.10
per query · USDC · Base · SKALE · Stellar
  • Multi-step deep analysis
  • Deep-research engine
  • Comprehensive source verification
  • Higher confidence scoring
  • Extended context window
View Manifest
For agent orchestration platforms

Platform-tier pricing — integrating verification at scale

Building a platform that needs verification as a primitive? Volume pricing available for high-throughput integrations. Includes white-label receipt branding, custom mapping documents under your namespace, priority support, and a dedicated integration channel.

Open-source verifier on PyPI: pip install agentoracle-receipt-verify — your end users get the same offline cryptographic guarantees regardless of which platform issued the receipt.

Talk to us View verifier on GitHub
technical

Specs & infrastructure

Protocolx402
ChainsBase · SKALE · Stellar
CurrencyUSDC + EURC
FacilitatorCoinbase CDP
Research EngineMulti-model consensus
Endpoints/preview (free) · /evaluate (free beta) · /research ($0.02) · /deep-research ($0.10)
HostingVercel Edge
Gasless OptionSKALE Active
Verification Sources4 (parallel)
Rate Limit100 req/hr
SDKLangChain · CrewAI
MCP ServerLive
faq

Common questions

Everything you need to know before integrating.

x402 is an open protocol by Coinbase that enables HTTP-native payments. Instead of API keys, your agent includes a payment proof in the X-PAYMENT header of each request. The server verifies the on-chain payment and returns the response.
Yes, your agent needs a wallet with USDC on Base. The /preview endpoint is completely free and requires no wallet — use it to test first. Most agent frameworks support wallet integration through Coinbase CDP.
USDC on Base (L2), USDC.e on SKALE Base (zero gas fees), and native USDC on Stellar. Same endpoint — agent picks the cheapest chain. EURC is also accepted on Base.
/evaluate (free during beta) verifies claims you already have — input any text, get per-claim ACT/VERIFY/REJECT verdicts. /research ($0.02) fetches new information from the web and returns structured results with confidence scoring.
100 requests per hour per IP for paid endpoints. 20 requests per hour for the free /preview endpoint. Rate limit headers are included on every response.

Your first verification is free

No wallet required. No API key sign-up. No payment method needed. Just paste a claim and verify.