@0x402/og

@0x402/og

0G Compute clients for TEE-attested inference, behind one interface.

npm i @0x402/og

import { createInferenceClient } from "@0x402/og";

const client = createInferenceClient({ kind: "router", apiKey, network: "testnet" });

const result = await client.audit({ token, artifact });

result.output;       // { action, maxLtvBps, findings, reasoning }
result.trustMode;    // "verified" when the provider ran it in a TEE
result.attestation;  // chatId, teeVerified, the signed response hash — or null
result.promptHash;   // what a seal commits to, so the claim is checkable later

Two backends behind one InferenceClient: the Router (kind: "router"), which asks for attested execution and carries the attestation back, and a direct provider client (kind: "direct").

Why the attestation is the point

An inference result nobody can vouch for is a sentence. One that comes back with teeVerified and a signed response hash is something an underwriter can be made to answer for — which is why @0x402/seal refuses a seal that claims an attested tier and carries no attestation, and why promptHash and responseHash are bound into the seal rather than left as prose.

The artifact is untrusted

artifact is attacker-authored material about a token — name, symbol, source, comments. It is never interpolated as instructions; see prompt.ts. Nothing downstream branches on output.reasoning either: it is model-authored text derived from that same untrusted input, and it exists to be read by a human, not acted on.

---

Repository · Site · MIT