@neuracomputes/agent — the agent-native SDK for NEURA Compute

OpenAI-compatible inference with the two things agents actually need:

1. Automatic HTTP 402 handling — when NEURA says payment required, the SDK reads the machine-readable remediation, burns $NEURA on Base, waits for the credit, and retries. No human, no card.
2. Wallet-only onboarding (SIWE / EIP-4361) — an agent holding only a wallet key creates its account and self-provisions an API key.

Use with the official OpenAI SDK (drop-in fetch)

import OpenAI from "openai";
import { NeuraAgent } from "@neuracomputes/agent";

const neura = new NeuraAgent({
  apiKey: process.env.NEURA_API_KEY,
  walletKey: process.env.AGENT_WALLET_KEY, // optional — enables self-funding
  autoFund: true,                          // burn->credit->retry on 402
  maxNeuraPerBurn: 100,                    // hard safety cap per burn
});

const openai = new OpenAI({
  baseURL: "https://app.neuracomputes.com/v1",
  apiKey: process.env.NEURA_API_KEY,
  fetch: neura.fetch, // 402s handled transparently
});

const r = await openai.chat.completions.create({
  model: "deepseek-ai/DeepSeek-V4-Pro",
  messages: [{ role: "user", content: "Hello!" }],
});

Zero-to-inference with only a wallet

const agent = new NeuraAgent({ walletKey: process.env.AGENT_WALLET_KEY, autoFund: true });

await agent.loginWithWallet();       // SIWE -> tenant + API key (no email, no form)
const reply = await agent.chat({     // 402 -> burnWithReason -> credited -> retry
  model: "deepseek-ai/DeepSeek-V4-Pro",
  messages: [{ role: "user", content: "You are funded by your own burn." }],
});

Safety model (burns are permanent)

• Nothing on-chain happens unless you pass walletKey and autoFund: true (or call fund() yourself).
• Every burn is sized from the live advertised rate; if no rate is advertised the SDK throws instead of guessing.
• maxNeuraPerBurn (default 100) hard-caps any single burn.
• On 402 without auto-funding, PaymentRequiredError carries the full remediation block and the x402 envelope so your own logic can decide.
• Use a dedicated hot wallet with a small balance — never a treasury key.

API

| Member | Purpose | |---|---| | new NeuraAgent(opts) | apiKey, baseUrl, walletKey, autoFund, maxNeuraPerBurn, creditsPerFund, burnConfirmations, rpcUrl | | agent.fetch | 402-aware fetch — pass to the OpenAI SDK | | agent.chat(params) | Plain chat completion with the same 402 handling | | agent.loginWithWallet() | SIWE login → { tenantId, address, apiKey } | | agent.fund(remediation, credits?) | Execute a burn + submit for crediting | | agent.creditBalance() | Current credits (1 = $1) | | agent.discovery() | /.well-known/neura.json | | PaymentRequiredError | Thrown on unhandled 402 — carries remediation + x402 |

Develop / publish

cd packages/neura-agent
npm install
npm test           # vitest unit tests (no network, no chain)
npm run build
npm publish --access public   # needs the @neura npm org