Following up on ARP, the reputation layer I’ve been building toward and Does it respond, is it maintained?. The loop described there now runs end to end.
ARP’s usage loop is live
Agents install skills, tools and MCP servers for us every day. ARP lets them check a skill before installing it, compare it with others that do the same job, and leave a line on Intuition once they have used it. The loop runs on Intuition testnet and adds to Claude in two minutes.
the problem
An agent picks a skill from stars, install counts and READMEs. None of them say what the skill will ask of you, or whether it did the job for anyone. When a skill fails, because a key is missing, a program isn’t installed or it was the wrong tool, the session ends and the next agent hits the same wall.
try it
In Claude Code, with Node 22:
claude plugin marketplace add intuition-box/ARP
claude plugin install arp@arp
In Claude, add a custom connector with this URL:
https://mcp.arp.intuition.box/mcp
Every other runtime is on the site: https://arp.intuition.box/
Name ARP when you ask: “find me a skill for cutting silences with ARP”. No wallet and no account needed to check a skill.
check a skill by its code
Before installing, the agent reads each skill’s files and shows what they contain: the commands they run, a login, the hosts they call, signs of a paid service. It shows the repository’s last push and whether it is archived. Tell it what you cannot provide, and it points out the skills that ask for it.
compare instead of choosing blind
The agent lines up the skills that do the same job, with what other agents recorded when they used each one, and in which domain.
after use, one line on Intuition
In Claude Code, the hook asks after each skill whether it did its job, and you choose once whether to share. A shared result becomes a chain of triples about the exact version of the skill that ran:
V = (cut-silences, version, sha256:<digest of the skill's files>)
T1 = (agent, uses, V)
T2 = (T1, in context of, video)
T3 = (T2, has outcome, failed)
T4 = (T3, because, missing-key)
T5 = (T4, needs, OPENAI_API_KEY)
The version is the digest of the skill’s directory. New files, new version, starting from zero, and every version rolls up to its skill. When the agent fetched the skill through ARP, the outcome carries a signed receipt for those exact files: (T3, has evidence, content-receipt:<hash>).
The reason comes from a short closed list. A failure for something missing on your side does not count against the skill: it records what the skill needs. A name like OPENAI_API_KEY is written only when the skill’s own SKILL.md names it. ARP’s sponsoring service pays the fees.
how a line becomes a rank
Each skill is scored per domain. Each wallet counts once, and the amount deposited is never read, so a bigger deposit buys no rank. A result moves the score once three distinct wallets say the same thing. Every line is on Intuition, so anyone can recompute the score.
beyond skills: paid calls and agents
Skills come first, but the same line already covers paid calls. When an agent pays an x402 endpoint and gets nothing back, it writes no-delivery, and a fifth level points to the settlement: (T3, paid with, eip155:<chain>/tx/<hash>). Those outcomes are published as an open feed, per endpoint. ARP also walks every ERC-8004 agent across chains, step by step: has a file, declares an address, answers, speaks MCP or A2A. That sieve is a feed too.
We took this to the people who work on these problems. A team that checks x402 endpoints before payment confirmed the gap sits after payment, where nobody keeps a shared record, and now has our feed. A team that makes paid calls for agents told us a per-endpoint record of paid and blocked calls would help them, with one rule we already follow: a failure on the buyer’s side never counts against the seller. An agent registry is looking at reading ARP’s lines in its own scoring. And a team building an index on Intuition will take ARP’s ontology as its first signal generator.
what’s coming next
A signal spec for readers. The ontology written down predicate by predicate, so an indexer or a scoring service can read these triples directly and weigh them its own way.
where you come in
If you build with agents, try it for a few days and tell me what helped, what got in the way and what’s missing. If you index Intuition or run a registry, the lines are also published as open feeds on the site: tell me in what shape you would read them. Replies here, or on GitHub: GitHub - intuition-box/ARP: ARP (Agent Reputation Protocol) — Monorepo bridging ERC-8004 agent identity with Intuition's semantic graph to deliver domain-modular, weighted reputation for autonomous agents. · GitHub