Documentation

Clyveris is one product with two surfaces. The desk is a reading experience for people. The agent is a paid research service for other software. Both run on the same corpus and the same rule: every claim traces back to a real, linkable source, and opinion is always labelled as opinion.

The desk

The dashboard lists the current signals. Each signal has a category, a publisher, and a detail page with three separated sections: what the source says (verbatim facts with a link to the original), the Clyveris take (labelled opinion), and one decision question to close the loop in a team conversation.

Filters narrow by category. The bookmark on each row saves a signal for your current browser session, saves stay on your device and are never sent anywhere.

Hiring the agent

Clyveris is listed on the CROO Agent Store as Clyveris Research Brief. Any CAP-capable agent (or a person using the Store's order flow) can send it a brief. Price is set on the listing, currently $0.10 USDC per brief with a 30 minute SLA. Payment escrows on Base and settles on-chain, gas is sponsored by the protocol.

The brief

Requirements are a JSON object. topic is required, 3 to 500 characters. Up to 10 tags sharpen matching. maxSources caps the returned sources between 1 and 10, defaulting to 3.

{
  "topic": "What is the cost of delaying a decision?",
  "tags": ["decision-making"],
  "maxSources": 3
}

Invalid briefs are rejected at negotiation with a reason, before any payment happens.

The deliverable

The result is structured JSON. status is either covered or no_coverage. Every source carries its publisher, URL, publication date, and the verbatim facts used, so the buyer can verify everything independently.

{
  "brief": "What is the cost of delaying a decision?",
  "status": "covered",
  "sources": [
    {
      "publisher": "Gartner",
      "url": "https://www.gartner.com/en/documents/7363830",
      "publishedAt": "2026-01-26",
      "facts": ["Decision intelligence brings decision modelling, analytics, and AI into one discipline."]
    }
  ],
  "editorialTake": "The hidden cost is rarely a single missed moment...",
  "decisionQuestion": "Which decision are we delaying because the context feels incomplete?",
  "tags": ["Decision-making"],
  "generatedAt": "2026-07-10T12:00:00.000Z"
}

Order lifecycle

Delivery is gated by a strict state machine. Nothing ships before payment confirms on-chain, and a delivered order can never be retroactively marked failed.

requested -> payment_required -> paid -> researching -> delivered
          \-> rejected (invalid brief)        \-> failed (expiry, payment failure)

A no_coverage result is a completed delivery: the buyer paid for an honest research pass, and honesty is the product.

This lifecycle has run for real. The first paid order settled on Base on July 10, 2026, and every step has a public receipt: escrow lock, delivery, and clearing.

Editorial principles

• Never fabricate a source. No verified match means no citation, stated plainly.
• Facts and opinion live in separate fields, in the data model and in the interface.
• Every source keeps its publisher, link, and date, provenance survives the whole pipeline.
• Popularity is a prompt for investigation, never evidence by itself.

Self-hosting

The whole product is MIT licensed. The README covers local setup in two commands, and docs/DEPLOYMENT.md walks through deploying the frontend to Vercel and the agent as an always-on service.