zhgg

zhgg is an on-chain autonomous agent commerce platform where AI agents don't just run in the cloud — they live on a blockchain, own wallets, earn yield, sign contracts, and pay for compute with real USDC, all without human intervention.

Each agent is an iNFT: an ERC-7857 token backed by a smart-contract receiver wallet (AgentReceiverWallet). The wallet parks idle balances into an ERC-4626 yield vault so capital is never sitting still. When an agent needs to act — audit another agent, hire a workflow, swap tokens, transfer funds — it does so through a full ERC stack: ERC-7710 for delegated spend-cap authorization so one agent can authorize another to spend on its behalf, and ERC-4337 UserOps routed through Pimlico's bundler for gasless account-abstraction execution.

The operator console is a terminal UI. Agents live in terminals — text streams, log tails, wallet addresses. A TUI is the natural habitat. The operator types natural-language intents: audit 2 mica, kh hire ml-risk-scorer, park 10 USDC, aa send 0xABCD... 0.01. The dispatcher translates each into signed transactions, UserOps, or REST calls in real time, and an audit trail panel streams every on-chain event as it lands.

The oracle agent runs TEE-backed inference on 0G Galileo, scoring agents against EU AI Act, MiCA, GDPR-AI, or price-oracle topics. Every score is committed on-chain via AxiomCommit (commitPlan → revealPlan) — a hash-locked audit trail that makes each agent's compliance history tamper-evident and publicly verifiable.

The KeeperHub integration closes the agentic-commerce loop: agents can discover third-party ML workflows on the KeeperHub marketplace, inspect their inputSchema and price, and autonomously hire them via x402 micropayments — no human in the loop, no API key shared with the caller.

Runtime stack: TypeScript throughout, running on Bun for sub-100ms cold starts and native ESM. The TUI uses raw ANSI escape codes and process.stdout.write — no ncurses, no blessed, no React. Each frame is built as a string diff and flushed atomically so there's never a partial-render flicker. A keypress handler translates operator input into typed IntentCommand unions that flow through a parser dispatcher to individual agent modules.

Chain integrations: All on-chain writes use viem. 0G Galileo (chainId 16602) hosts the AgentNFT registry, AxiomCommit plan hashing, and AgentReceiverWallet yield vaults (ERC-4626). Base Sepolia hosts the DelegationManager for ERC-7710 spend-cap delegations. Pimlico's ERC-4337 v0.7 bundler handles UserOp submission and sponsorship for the AA wallet slice.

The hacky part — 0G receipt polling: 0G Galileo's testnet nodes reject eth_getTransactionReceipt with an immediate error (not a null response) for any pending transaction. viem's waitForTransactionReceipt interprets that first-poll error as a hard failure and throws immediately, well before the configured timeout. The fix: fire-and-forget. We submit the transaction, snapshot the block number before submission (needed to deterministically derive the AxiomCommit commitId via keccak256(tokenId ++ planHash ++ sender ++ blockNumber)), return the derived ID to the caller immediately, and let waitForTransactionReceipt confirm in the background — errors silently swallowed as non-fatal. The on-chain record lands; the UX never blocks.

The hacky part — allowance: ERC-20 approve(amount) on a re-used demo wallet gets consumed after the first run, causing ERC20: transfer amount exceeds allowance on every subsequent orchestration. Fix: check current allowance before every split; only re-approve if it's already below the payment amount, and approve maxUint256 so the approval survives indefinitely across runs.

KeeperHub (partner tech): The kh discover / kh inspect / kh hire intent chain lets the TUI browse KeeperHub's public MCP workflow marketplace (~85 entries), validate required inputSchema keys client-side, and pay-and-invoke via x402 in a single operator command. The bearer key lives only in the agent process; it's never surfaced in the TUI or commit history.

Oracle + TEE (0G): The oracle agent sends a compliance-check prompt to a TEE-backed inference endpoint on 0G, receives a structured score, records the result on-chain via giveFeedback (ERC-8004 pattern), and commits the plan hash via AxiomCommit — all within a single audit intent. Topics (eu-ai-act, mica, gdpr-ai, price) are parsed at the TUI intent layer and passed through the full dispatch chain typed.