Tanilo runs model checks on submitted claims and, when signing succeeds, returns a signed receipt. Its gate decision can be recomputed from the recorded gate inputs using the referenced published mapping. This does not reproduce the model judgments or establish that the claim is true.
Verification over the open world is not deterministic — the judgement is model inference, and we say so rather than claiming otherwise.
What we make deterministic is the derivation: a published rule and canonical bytes. Given the receipt and the specification, a third party recomputes the gate decision offline, without calling us and without trusting us.
One API call sits between your agent and its consequential actions. Here is the whole product:
"This invoice is unpaid." "This drug interaction is safe." "This customer is eligible." The claim that justifies the action.
Model checks, including one that argues against the claim, combined under a published rule — before the action runs, not after it goes wrong.
Canonical bytes (RFC 8785), Ed25519 signature, the verdict, the recorded gate inputs and the referenced mapping — sealed at issue, when signing succeeds.
Offline, against published keys. Auditors, counterparties, courts — no account needed, no trust in us required.
Don't take the homepage's word for anything. Pull a real receipt and verify it on your own machine, offline.
The verifier is MIT-licensed, no dependencies on us. Read it before you trust it — that's the point.
This receipt is signed with ao-composed-2026-06-ed25519-c3abfce3. Until 28 August 2026 that key could also sign bytes supplied by anyone, so the signature alone proves the bytes haven't changed since signing, not that we issued them; what corroborates issuance is a check against our internal claim-fingerprint store, which we run on request.
agentoracle-mcp is a Model Context Protocol server for Cursor, Claude Desktop, Codex, and any MCP-compatible client. Verify signed receipts, check confidence scores, and pull the current JWKS — from inside your editor.
Point at a receipt file or URL, get valid: True — or the exact reason it isn't — without leaving the editor.
The MCP server runs locally, verifies against the public JWKS, and never phones home.
Receipts the MCP verifier checks are byte-identical to the ones the reference verifiers produce.
Published on npm as agentoracle-mcp.
A log is the operator's story. A receipt carries a decision a third party recomputes — verdict, rules, signature.
Letting the operator keep the record is letting the suspect write the police report.
Change one byte and the signature fails. Backdating fails the same way — timestamps are bound in.
Published keys, canonical bytes. If we disappeared tomorrow, every receipt still verifies.
A separate team, working from the published specification text alone and without access to our implementation, rebuilt the format — matching to the last byte.
IETF-filed, MIT reference code, published conformance vectors in two languages.
The EU AI Act's record-keeping obligations (Article 12) apply from 2 December 2027 for standalone Annex III systems and 2 August 2028 for AI embedded in Annex I products, under Regulation (EU) 2026/1744. Receipts are records an external examiner can verify without trusting you — built for exactly that clause.
Give every consequential agent action an evidence trail. One API call in the loop; a receipt in every audit bundle, ready for whatever governance stack sits above you.
Finance, health, legal — anywhere "the AI said so" isn't a defense. When something is questioned, hand over proof instead of promises.
Pricing and payment details on this page are out of date. Current information: https://tanilo.io/pricing
Partner in the Hashgraph Online Partner Program. Participant in the Cryptographic Agent Execution working group. Partner profile ↗
Pricing and payment details on this page are out of date. Current information: https://tanilo.io/pricing
The signed fields record which claim was evaluated (by its hash), when, under which rule table, and the verdict; changing any of them breaks the signature. The signature doesn't independently prove how the checks ran. Sources listed with the result are for reference: they aren't part of the signed receipt and don't feed the verdict yet.
Install the MIT-licensed verifier (or write your own from the IETF draft — a team with no relationship to us already has, byte-identically). Verification runs offline against our published public keys. You never need our permission, our API, or our continued existence.
Article 12 requires records of high-risk AI operation, applicable from 2 December 2027 for standalone Annex III systems and 2 August 2028 for AI embedded in Annex I products, under Regulation (EU) 2026/1744. Receipts are records that verify independently — the property plain logs can't offer an examiner. We publish a free Article 12 mapping in the whitepaper.
When signing succeeds, you get a signed receipt whose recorded gate decision is halt (shown as do_not_act on this page), with the result behind it in the signed fields (for example v_verdict refuted or unverifiable, or v_adversarial_result vulnerable — not surviving the counter-argument check). The "no" is evidence too — often the more valuable kind.