Assay gives every open-weight AI response a receipt. It names the model and the host, the host signs it and anchors it on Monad, and you can co-sign it with your passkey.
It proves who served which bytes and what they claimed. Not which weights ran.
Your prompt and answer never go onchain. Only salted hashes do, and you keep the salt.
An animation of dots: a printed receipt becomes a Merkle batch, then a passkey fingerprint, then the Assay mark.
This receipt lives on mainnet. A receipt never moves between chains, so it looks the same on either network.
This is the first receipt Assay anchored on Monad mainnet, on 5 Oct 2026. Every value on it is real. Scroll, and each part is stamped as it's explained.
This receipt proves which host served these bytes and what it claimed. It does not prove which weights ran. Your prompt and answer never go onchain. Only salted hashes do.
KEEP YOUR SALTassay-ten-xi.vercel.app
Read from the host's public API, 11 Oct 2026
Host 10278 signed these exact bytes with an ES256 key tied to its ERC-8004 identity. Your browser checks the signature against the host's published key.
gemma-4-31b-it sits inside the signed body. If the claim were false, it would be a signed false claim the host can't take back.
The receipt sits in a batch whose root landed in block 110,684,366. Monad's P256 precompile checked the host's signature on that root.
Nobody co-signed this one, so its YOU stamp stays empty. When you ask through Assay, your passkey can add the fourth mark. No wallet needed.
It proves who served which bytes and what they claimed. It doesn't prove which weights ran. That gap is why hosts also get graded against the lab's own endpoint.
Open any receipt on its own page: the host's signature, its batch on Monad, the host's grade from verifiers you trust, and a line that reproduces each check.
The text you get back from an open-weight model carries no proof of who served it, or of what they did with your request.
I asked 16 hosts serving Kimi K2.6 for at most 16 tokens. Thirteen stopped at 16. Three kept going in 59 of 60 tries, and billed up to 126 times what Moonshot's own endpoint charged.
One model, one prompt, one day: 8 Oct 2026, 350 requests. Hosts are named on 25 Oct, after their 7 days to reply. Read the write-up
endpoints serving one open-weight model, from different providers.
of them don't state their precision at all.
declare a 4-bit format.
per million output tokens, cheapest endpoint to priciest.
DeepSeek V4 Flash on GPQA through the lab's own API, and through one third-party host. Measured by Olly.
of 735 endpoints across 140 open-weight models list their precision as unknown.
Source: the Assay hosting report v0, built from OpenRouter's public endpoints API on 26 Sep 2026. Read the report
Open the real thing. These values were read from Monad mainnet and the Envio indexer on 11 Oct 2026.
Kanmani's USDC escrow on Monad bought 5 answers from host 10278. It published the Merkle root of the receipt hashes in advance, both sides checked each receipt against the anchored batches, and then the escrow settled.
Every check runs in your browser.
Open it Hosts on mainnetTwo hosts: Gemma 4 31B, and Kimi K2.6 from Moonshot's own endpoint.
Open a host Grades on mainnetAgainst Moonshot's own endpoint, with the hash of the raw logs posted onchain.
Look up a grade Built on by other teamsThat gap is why hosts get graded against the lab's own endpoint, with the raw logs published. Hosts that sign from inside a TEE, with a model hash in the attestation, are next on the roadmap.
Read the threat modelThe host signed what it served and the batch is on Monad.
Open verifiers test the host against the lab's own endpoint.
A deterministic runtime lets anyone re-run a revealed request.
The host signs from an enclave whose attestation names the model.
The SDK sends the salt and your co-signer key with each request, parses the receipt from the response and checks that the output commit matches what you received.
$ npm install assay-receiptsFrom an empty folder, it verifies the first mainnet receipt in about a second.
$ npm install assay-receipts + assay-receipts@0.1.1 $ npx tsx verify.ts 0x1b443b45...12e35c5f2 jws pass, hash pass, kid pass, merkle pass, anchored pass, outputCommit skipped, promptCommit skipped, cosigned skipped. Skipped means the check needs your salt, or the receipt names no co-signer.
import { wrap } from "assay-receipts";
const ask = wrap(fetch, { cosigner }); // your passkey's key hash
const { json, receipt, salt } = await ask(
`${HOST}/v1/chat/completions`,
{ method: "POST", body: JSON.stringify({ model, messages }) },
);
// receipt.hash lands in a Merkle batch on Monad.
// Keep salt: it opens the commits later.import { verifyReceipt } from "assay-receipts";
const hash = process.argv[2];
// The receipt, its batch proof and root, from the host that served it.
const { body, jws, proof, root } = await fetch(`${HOST}/v1/receipts/${hash}`).then((r) => r.json());
const jwks = await fetch(`${HOST}/.well-known/jwks.json`).then((r) => r.json());
// client: any viem PublicClient on Monad. anchor: ReceiptAnchor on mainnet.
const anchor = "0x049A73755cA3508ef3Daa4752A3406f6e00CfB13";
const { checks } = await verifyReceipt({ body, jws, jwks, proof, root, onchain: { client, anchor } });
for (const [name, r] of Object.entries(checks)) console.log(name.padEnd(13), r);Paste a receipt, and the salt and output if you kept them. Every check runs in your browser or as a free onchain read. No wallet needed.