Lepton Agents Hackathon (Canteen × Circle × Arc): agent-native nanopayments on Arc, built on the existing SettleKit platform. Hackathon modules (demo-ready, RFB-aligned): - @settlekit/agent: autonomous paying agent (Claude tool-use loop) — RFB 1/3 - @settlekit/citation-toll: per-access content + recursive royalty splits — RFB 6 - @settlekit/streaming: per-second continuous-authorization settlement — RFB 4 - @settlekit/x402-client: x402 pay-and-retry + pluggable settler (local/Circle) - @settlekit/agent-economy: closed-loop traction harness + CLI - apps/api /v1/lepton: public live demo surface Production settlement spine (Phase 1): - @settlekit/settlement-core: configurable Gateway/Circle/local providers, idempotency, batch accumulator, on-chain reconciliation, nonce + receipt stores - @settlekit/wallet-fleet: registry, spending caps + kill-switch, balances, Circle provisioning - x402-client: provider-settler adapter + Arc-indexer on-chain verifier - lepton DB schema (12 tables) + persistence Pg stores (settlements, nonces, sources, royalty legs, wallets, streams, payees) - apps/worker jobs: lepton-settlement-reconcile, lepton-payout-sweep, lepton-stream-refund (no-op until the settlement spine is wired) First real-world traction sidecar (Phase 2): - @settlekit/payee-registry: identity→wallet payout rule (in-memory + Pg) - services/sidecars/rsshub-citation-toll: ingest RSS items as priced citeable sources, x402-gated, settle per-citation royalties via settlement-core - shared sweepPendingRoyalties used by the worker job and the sidecar - services/sidecars/* wired as a TS workspace area (pnpm + vitest) Docs: LEPTON.md, LEPTON_PRODUCTION_PLAN.md. All green: full typecheck + 870 tests. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
1.4 KiB
1.4 KiB
@settlekit/agent
An autonomous paying agent for nanopayment marketplaces (Lepton RFB 1 & 3).
It discovers x402-priced services, decides which are worth paying for under a hard spend policy, settles the toll via @settlekit/x402-client, consumes the result, and rates the service. The decision engine is pluggable:
ClaudeDecisionEngine— the real brain. Exposes the capability surface as tools to Claude (claude-opus-4-8) through the official Anthropic SDK tool runner; the model drives the trajectory. RequiresANTHROPIC_API_KEY.HeuristicDecisionEngine— deterministic fallback for offline runs, CI, and the closed-loop harness.
Every spend guardrail (total budget, per-call cap, reputation floor, purchase limit) is enforced inside the capability layer, so no engine can exceed the policy regardless of what it decides.
import { PayingAgent, ClaudeDecisionEngine } from "@settlekit/agent";
const agent = new PayingAgent({
services, reputation, fetcher, settler, from: "0xagent",
policy: { totalBudgetUsdc: "0.01", maxPriceUsdc: "0.002", maxPurchases: 5, minReputation: 3 },
});
const result = await agent.run(
{ objective: "Gather grounded sources on Arc settlement within budget." },
new ClaudeDecisionEngine(), // or omit to auto-pick by ANTHROPIC_API_KEY
);
// result.purchases, result.totalSpent, result.remaining, result.log