Keryx 🏛️

Keryx is a citation-toll reading agent for researchers and builders: a question and budget lead to a cited report, visible BUY/SKIP/CACHE decisions, and payment receipts for eligible creator rewards.

Guided recorded walkthrough · Live app · Public proof · Archived QA walkthrough (2m40s, app 0.27.43, recorded Oct 8). Reopens an archived first-party report and receipt; no new research was submitted. See the recording scope and public checks.

Try it in two minutes — no wallet

1. Open this existing English cited report about how source-access tolls differ from citation rewards. Read the summary and its [S1] source excerpts, including the Low confidence boundary.
2. Expand Decision log below the answer. Inspect why Keryx reused the cached first-party article and skipped unrelated candidates. This run has CACHE/SKIP decisions; it did not pay a new access toll.
3. Open View JSON in Portable research receipt, or use the direct receipt. Under payload.settlement, check mode: real, ledgerCompleteness: complete, 0 USDC access / 0.025 USDC citation, and the mainnet Circle reference. Receipt verification and the retained digest explain what these records prove.

This is an archived owner-operated QA result using Keryx Engineering (first-party), not a fresh answer or an outside customer's task. The location/job in its prompt is a test persona. It supports the three core payment facts but omits the requested document revision, so the original deliverable remains partial. Opening the report or receipt makes no research or payment request.

Next, ask your own question: open keryx.cc, choose Quick, and submit a shareable question. The public sponsored trial needs no wallet or sign-up; the composer shows the payer and budget. Admission depends on current quota, available capacity and spend caps, so a new answer within two minutes is not guaranteed. Submitted questions become public reports. See sponsored trial limits.

Usage so far

Read-only snapshot of public metrics, October 8, 2026, 16:24 UTC, with public health reporting Arc mainnet / real settlement at 3b839ccd.

Measure	Observed value	Counting boundary
Recorded research runs	50	All caller origins, including team and QA; not accepted deliverables
Settled payment records	7	All payment kinds; pending, failed and simulated excluded
Settled volume	0.095 USDC	Gross payment volume, not net revenue
Creator access + citation payouts	0.065 USDC to 2 earning creators	Includes first-party recipients; not independent creator adoption
Settled operating fees	0 USDC	Reported separately from creator rewards
Outside users / team / QA	Separate counts unavailable	The endpoint does not partition these groups; guest questions are not unique outside users
Historical Arc testnet	Separate evidence only	Not included in these mainnet totals; see the event-period record

These are dated database projections. Live Ledger and Public Proof expose the current records and their limits. The submission pack separately documents sponsored outside use on testnet and unmatched participant stories; none is inferred from these aggregate counts. Snapshot hashes retain the exact observation.

What is real and what is scripted

• The two-minute path reopens a stored run. Its model-written answer, decision trace and receipt are recorded artifacts. It is a curated QA example, not a live rerun or evidence of customer acceptance.
• The example's payment is recorded as real Arc mainnet settlement: one 0.025 USDC citation reward to the first-party Engineering source. Its cached read cost zero. The Circle transfer reference is not a per-citation Arc transaction hash.
• First-party and public sources are distinct. Keryx Engineering and Keryx owner release notes are first-party validation sources. Public web/paper references alone grant no creator payout rights. Historical seeded sample sources have Keryx-controlled authority; they do not establish independent seller adoption.
• CLI demo, seed and benchmark commands are scripts. Offline development uses labelled simulated payments. Invoking a real-mode script can make actual paid requests; it is not required for the wallet-free review path.
• Production research is caller-driven. The local desktop alpha is not an autonomous scheduler. Monitoring/reconciliation jobs are operational checks, not independent users or fresh research. This public checkpoint does not verify the host's current schedule inventory.
• Settlement is checked per receipt. Pending, failed, simulated and planned amounts are separate from settled value. A cited answer, healthy API or script completion alone does not prove a payment settled.

What happened during Tameion

The pre-event main baseline is 2291753 (September 25). During the event Keryx added original-document and scholarly discovery, portable cited-report exports, a local CLI/Windows task workspace, bounded manual Research Monthly, and prepaid Operator recovery/observation. Production moved from Arc testnet to mainnet on October 4. The event-period product delta links source evidence and the limits of each change.

The usage record separates the event's testnet phase from mainnet; the mainnet snapshot above includes internal activity. Sponsored outside use is documented, but independent paid demand, repeat use and general deliverable usefulness remain separate evidence gates. Source candidates beyond the observed deployment are not claimed live. The dated archived-QA walkthrough above reopens existing artifacts; it is not a live-Ask or complete current-feature demonstration.

---

The problem

The web's economics assume a human reader: you write, people visit, attention becomes ads, subscriptions, tips. AI agents broke that contract. They read everything and send back nothing — no click, no view, no cent. Every answer an assistant gives is built on someone's work, and that someone is invisible at the exact moment their work proves its value.

Keryx (κῆρυξ — herald) exists to fix that one moment. It makes citation itself the payment event: the instant an agent relies on your writing to answer a question, you are paid, proportional to how much you helped, settled sub-cent in USDC on Arc. No accounts, no invoices, no ad tech. A creator onboards by pasting an RSS URL. An agent pays because paying is cheaper than not knowing.

What Keryx is

Keryx is a citation-toll reading agent — an autonomous reader with a wallet, and the payment rail underneath it. Give it a question and a budget:

1. It decomposes the question into sub-claims and discovers exact article candidates from free previews and the signed offer book.
2. It decides, per article version: buy / skip / cache — expected value vs effective price vs remaining budget, with a human-readable rationale for every choice.
3. It pays the x402 toll only for what it buys, checks sufficiency after each read, and stops early when it has enough.
4. It returns a summary cited sentence by sentence: each sentence is written from one verbatim source excerpt, checked against it, and shown beside it, with explicitly labelled research targets/gaps. Uncited prose is withheld; the answer retains a Low/incomplete boundary, and allocates bounded citation rewards only to eligible accepted sources. Creator payouts are settled only when recorded payment evidence confirms settlement; multi-author rewards follow their payout split.

Ordinary short-bullet requests can retain their requested count and supported output language when every target, reviewed sentence and qualifying excerpt fits the same grounded layout. Otherwise the answer keeps inspectable target sections and explains the unmet format. See candidate scope and acceptance. Unsupported output-language requests use English scaffold labels without forcing statement translation; language-selection scope records the finite parser and remaining localization gates.

The result is a working micro-economy: readers that pay by default, and writers that earn by being useful — not by being clicked.

An agent that genuinely decides

Most "payment agents" are scripts with a wallet. Keryx's differentiator is visible agency — the model reasons about money and shows its work, streamed live to the UI:

• Buy / skip / cache with rationale — every spend decision names the exact article version and explains why.
• Open article market — /market and GET /api/offers publish exact payable versions, registry list prices, x402 paths, and verifiable EIP-712 discounts signed by publishers.
• Source comparisons: recorded exact excerpts let readers inspect differing sources. The current claim-grounding candidate does not certify an independent synthesized conclusion about which source is correct.
• Confidence boundary: the answer is labelled Low/incomplete. Qualified source excerpts provide recorded support; target coverage estimates do not verify each assertion or establish complete useful synthesis.
• Sentence-cited summaries: each summary sentence is written from one qualified literal source excerpt, scored against that excerpt by a separate model review, and rendered beside the verbatim source text. Sentences without a qualifying excerpt or review are removed, and the answer falls back to excerpts alone when none survives. The review is a model judgment, not proof; cross-source conclusions are not written. Accepted excerpt citations still support rewards, with original paid-fetch debits, article bindings and settlement identities. Production now uses Arc mainnet; see the current deployment and evidence limits. Deployment does not establish complete useful synthesis or independent acceptance.
• Evidence ledger — every rewarded citation carries a claim-indexed exact quote. The orchestrator verifies that quote against content it actually read before the marker can receive a citation reward; rejected markers are removed from the answer.
• Living answer receipts — archived conclusions never silently change. Keryx flags when an exact paid SHA-256/IPFS article version has been superseded, exposes a metadata-only freshness API, and itemizes the source/evidence/coverage/settled-payment delta after a reader explicitly re-asks.
• Researcher exports — import recorded cited-article references as BibTeX or RIS into Zotero, and compare claim-level excerpts in a research evidence matrix with spreadsheet CSV. Observed scholarly records include supplied authors/DOI/journal metadata and read limits; missing fields stay explicit. See researcher exports.
• DOI and scholarly discovery — resolve up to two exact Crossref DOIs from a question, or opt into Crossref/arXiv searches. Read selected original publisher pages and bounded versioned arXiv PDFs, with explicit abstract-only fallback. Metadata and public papers carry no creator payout authority. See scholarly research and the separate proposed author opt-in payment plan.
• Portable research receipts — every permalink exports one deterministic JSON bundle containing its answer hash, BUY/SKIP/CACHE decisions, exact article versions, claim evidence and sanitized Circle settlement snapshot. Retain its SHA-256 to detect later changes; the self-check does not pretend to be a Keryx, publisher, or on-chain signature.
• Claim-aware evidence portfolios — Keryx chooses a non-redundant set of positive source proposals under separate attention and fetch-USDC caps. CACHE correctly costs zero fetch USDC but one context slot; the receipt compares preview-predicted coverage with evidence yield after read.
• Cross-query memory — sources that proved useful (or useless) in past runs on the same subject shift future buy/skip decisions. A source is scored against the runs that actually read it, so a skip never becomes evidence against the source it skipped.
• Semantic discovery — candidate matching by embedding similarity, not keyword luck.
• Emergent frugality — it stops early, reuses its cache, and correctly spends nothing when nothing is worth buying.

Money safety is enforced in code, not by the model: the LLM proposes value; the orchestrator enforces the hard budget cap, so a hallucinated number can never overspend. An economic-invariant test suite (spend ≤ budget, payouts = weights, splits sum exactly) runs in CI on every push.

The controlled offline paying-source study runs the actual Keryx heuristic pipeline on four published fictional questions across four budgets and three catalogue prices. Retained bodies, actual synthetic outputs, literal/read-bound scoring and exact simulated ledger amounts are reproducible with npm run eval:paying-source-study -- --check. It measures this closed fixture only; live model behavior, semantic correctness, real spending and exploration's causal effect remain unmeasured.

Historical testnet trace (recorded output; not current mainnet traction):

[decide]  BUY Agent Economy Weekly — strong match on x402, autonomous, commerce; worth the $0.004 toll
[decide]  SKIP Garden & Soil Monthly — weak match (no key terms); not worth $0.002
[fetch]   Paid $0.004 to Agent Economy Weekly — S1
[sufficiency] Read 2 sources covering all sub-claims; stopping early to save budget
[settle]  Settled $0.015 → Mara Okoye · $0.010 → Devin Park   (60/40 author split)
📊 $0.032 spent → 100% to creators · 3 bought / 3 skipped

Research tools and desktop

Windows local Operator alpha: npm run desktop:install then npm run desktop:start. See the desktop guide for the unpacked app build, offline saved results, private Markdown briefs, and limits.

Research-paper library: browse 40 observed paper records from arXiv, OpenReview, PMLR and ACL Anthology, filter by author/year/DOI, or explicitly search live arXiv/Crossref metadata. /api/papers and npm run papers share the bibliography-only contract. Links remain unread until a separate research run.

Literature workspace: save exact paper records on this browser, screen against a review question, keep personal notes, export CSV/JSON/RIS and prepare an editable two-paper comparison. No account, cloud sync or automatic ask.

Research Monthly: a bounded plan for four Deep requests over 30 days, 10% below four separate packages with unchanged creator caps. Manual renewal; no scheduling or unlimited use. Failed/pending jobs use a slot. Web/API own the entitlement; Monthly CLI and MCP handoffs share that contract.

Original-customer deliverable choices source candidate: an ordinary settled prepaid A2A customer can Accept, request a revision or Reject the exact retained delivery through its report, scoped API, MCP or npm run deliverable:acceptance -- read <a2a_id>. Submissions use the exact read-back fingerprint/digest/revision and a retained idempotency key. Reasons stay private; sharing current choice/timestamp is opt-in and withdrawable. Historical remedy:none and missing response-window terms remain visible. Revision/refund execution is withheld; Monthly/private/browser/native originals and synchronized deployment/distribution acceptance remain separate gates.

For creators

• Claim proven demand — keryx.cc/wanted shows claims paid dispatches left under-covered. A new creator can check and list an RSS feed; an existing registry creator can offer the exact indexed article directly and optionally sign a temporary discount. Keryx guarantees that version a candidate slot, not a purchase, then queues one bounded retry after refreshing creator authority. Fulfillment is public only when that source passes the evidence gate and its citation reward really settles. Every claim has a canonical shareable brief and social card, so the specific gap can reach the writer who covers it.
• Onboard from your own wallet — paste an RSS feed at keryx.cc/register and publish your source to the on-chain registry. Keryx sets up the x402-priced endpoint and free preview; it never holds your key. On mainnet your wallet supplies native USDC gas; the faucet is available only on testnet.
• Own your payout — your registered source pays the wallet you signed in with. The first historically owner-verified testnet creator (conzit.com) proved feed ownership, set its address, was cited & paid end-to-end — and has since claimed its registry record from its own wallet, so its on-chain creator is the creator, not Keryx. We've also proposed this as an opt-in convention upstream in RSSHub. Seeded sample sources have Keryx-controlled registry authority; live authority and payout checks are available on /proof.
• Know the moment you're cited — opt into a plain email alert (no webhook server needed, rate-capped, one-click unsubscribe) and/or signed webhooks that fire the instant a citation settles; every payout on your public earnings page shows the actual question your work helped answer.
• Show it off — an embeddable "Cited by Keryx" badge (live SVG at /api/creator/<id>/badge.svg) displays your real citation count + USDC earned on your own site, with copy-paste Markdown/HTML on each creator page. Payouts become portable, verifiable proof.
• Cash out yourself — your wallet signs a Gateway burn intent in the browser and submits the reviewed mainnet mint with native USDC gas. Optional gas-relay operation is a separate role; historical testnet cash-outs do not prove a mainnet withdrawal.
• Keep everything — 100% of every citation reward goes to creator wallets. 0% platform fee.
• Squat-proof identity — sources live in an on-chain SourceRegistry (0x42a64061b6cd84067bb660b2a9b8aa881fd225bb) with creator-scoped IDs and on-chain multi-author splits.

Public sources and owner claims

Public web discovery finds supported websites, feeds and PDFs for the question at hand and reads available content for free. /sources shows retained feeds and creator listings, rather than a bulk index of the internet. A publisher can choose This is my source or open /claim-source, sign in with its wallet, and prove control using a dedicated website file or publisher-controlled RSS/Atom channel token.

Verification earns nothing by itself. Connect the exact owner-created registry listing and explicitly choose free reads with no rewards, zero-price reads with qualified citation rewards, or positive-price paid reads. Earning activation requires separate distribution-rights consent and fresh control proof; registry registration is an owner-reviewed network transaction with possible native gas, using no agent funds. Control expires after 24 hours and earning eligibility pauses until explicitly refreshed. Old public reference identities and free receipts remain free, and no activation bills past uses. Zero-price reads require no x402 access authorization or settlement receipt.

The shared reading pipeline applies the same policy and evidence gates across web, API and agent clients. Claim management uses the web and its authenticated API; scholarly-rights enrollment remains separately gated. See the public-source claim guide for proof, listing, policy and recovery steps and deployment capability requirements.

For developers & agents