internet-court/internet-court-skill

The trust problem in agent-to-agent commerce is not payments — it is what happens when a payment goes wrong. Internet Court is a single agent skill built around that gap. Payments, identity, negotiation, and execution already have protocol coverage; the missing piece has always been adjudication, and this skill makes it a first-class layer rather than an afterthought.

The architecture is a master router — a SKILL.md that pulls sub-skills into context on demand — sitting atop 91 vendored skills from 33 protocol owners. The vendored set covers six layers: discovery and identity via ERC-8004 and ERC-7857, negotiation via A2A, contracts via ERC-7710 and Arkhai's natural-language agreement tooling, payments via x402 and several multi-party payment protocols, execution across compute and data rails, and finally verification and disputes via GenLayer, Kleros, and UMA. The first five layers are largely assembled from existing published skills; the sixth is what Internet Court claims to own.

The dispute-resolution design is worth understanding precisely. When two agents form a contract, they select their arbitrator in advance — GenLayer, Kleros, UMA, or another — and that choice is embedded in the contract structure itself. The skill does not impose a single arbiter; it standardizes the slot where the choice lives. Most deals never reach adjudication. The point is that the mechanism exists before the deal closes, not after something breaks.

Twenty-eight founding consortium members span every layer of the stack, from MetaMask and OKX to Kleros and Arkhai. The README is candid that not every member has shipped a public skill yet — ERC-7857, ERC-8183, and UMA are noted as having no public skill available at the time of writing. The vendored table is the honest inventory of what actually exists today, and skills-lock.json pins each source with a hash and a refresh command.

Installation is deliberately broad: Claude Code plugin marketplace, npx skills, Codex, opencode, OpenClaw, and Hermes are all documented. The root-only install works via a raw-URL fallback to GitHub for sub-skills not bundled locally, which means a minimal install is functional but network-dependent.

The framing — that fragmented protocols each solve the happy path and pass failures down the line — is accurate and the design follows from it. Whether the consortium's open governance actually holds as member interests diverge is the real question this README cannot answer. But the structural bet is sound: accountability requires adjudication, and adjudication needs to be in the contract from the start, not bolted on after a dispute surfaces.