Skip to content

Releases: tang-vu/keryx

v0.4.0 - the agent gets judgment, creators get a value loop, ops get hardened

Choose a tag to compare

@tang-vu tang-vu released this 02 Jul 06:30

v0.4.0 — the agent gets judgment, creators get a value loop, ops get hardened

Agent brain:

  • Multi-pass reasoning with per-claim confidence sufficiency
  • Adjudicates conflicting sources instead of averaging them, with a logged rationale
  • Confidence verdict: the agent rates and hedges its own answer
  • Cross-query memory + semantic discovery via embedding similarity

Creator value loop:

  • Non-custodial creator cash-out from the app (Circle fee reserved before signing)
  • Feed-ownership verification gates citation earnings
  • Creator earnings pages + ERC-8004 reputation scoring
  • Notify-on-citation webhooks: signed POST the moment a source is cited and paid
  • Payouts show the triggering question; per-creator and per-dispatch social cards

Distribution and on-ramps:

  • keryx-mcp published to npm + the official MCP registry (npx one-liner)
  • One-call testnet onramp; obvious free no-wallet trial
  • One-command full-cycle demo (npm run demo) with honest on-chain proof
  • Shareable dispatch permalinks + prefill/auto-run ask links
  • Onboarding tour, answer-quality feedback, mobile navigation

Reliability and ops:

  • Economic-invariant vitest suite (22 tests) + CI typecheck/test gates
  • Exact micro-USDC multi-author split allocation (largest remainder)
  • Rotating SQLite backups with off-box copy; treasury watchdog + failed-settlement alerts
  • Low-downtime redeploy, /api/health + /status page, unbuffered live SSE trace

v0.3.0 — Visible agency + external-agent onboarding

Choose a tag to compare

@tang-vu tang-vu released this 21 Jun 21:22

Release wave since v0.2.0 (64 commits). All live at keryx.cc, real settled USDC on Arc testnet.

✨ New visible features

  • Live budget meter in the reasoning console — watch the agent spend against its authorized budget in real time as it buys sources and pays creators per citation, and see it stop early with USDC unspent. The cap is enforced in code, not by the model; now it's visible.
  • "Call Keryx from your own agent" card on the dashboard — copy-paste the exact x402 call (free inspect + circle services pay) so any agent can pay Keryx and have it research + pay the creators it cites downstream.

🔒 Hardening

  • Every public spend path capped + IP rate-limited: anonymous /api/ask (treasury) budget clamp, A2A /api/agent/ask budget clamp, /api/cite reward ceiling — the live treasury can't be drained or volume faked.
  • Demo-path hardening pass: settlement-loop guarded (one failed toll no longer kills a run), resilient LLM engine (retry → heuristic fallback), branded error/SSE-disconnect states.

🌐 Platform

  • Non-custodial browser co-sign session-payment path made reliable end-to-end (deposit → activate → pay → settle).
  • On-chain SourceRegistry catalog published with verifiable provenance (ArcScan-verified).
  • A2A endpoint discoverable — serves the x402 challenge on GET so circle services inspect reports it payable.
  • 24/7 traction daemon; honest external-vs-engine volume split on the dashboard.

Full detail: docs/project-changelog.md.

Keryx v0.2.0 — Open-economy reasoning

Choose a tag to compare

@tang-vu tang-vu released this 17 Jun 15:26

Keryx v0.2.0 — Open-economy reasoning

Highlights since v0.1.0:

  • Open x402 marketplace discovery: the agent probes the live Circle x402 service
    bazaar (circle services search) each query and reasons BUY/SKIP over real
    third-party endpoints alongside its own creators. Off-Arc endpoints are
    evaluated and logged but never settled (in-code rail constraint, like the
    budget cap). A "x402 market" badge marks them in the live reasoning trace.
  • "The Mint" visual system: full engraved-banknote redesign — Bodoni, guilloche,
    herald seal, coin/engraver iconography, cursor-steered hero globe.
  • Live dashboard real-data: real most-cited topics (source tags x citations) and
    a 12-week creator-earnings chart.
  • x402 settlement hardening: clear Circle's 7-day authorization validity floor
    with margin to stop intermittent verify failures.
  • Broader question bank (AI agents, Ethereum, stablecoin payments); deploy at
    keryx.cc via VPS + Cloudflare Tunnel (npm run deploy).

Keryx v0.1.0 — citation-toll reading agent (Lepton Hackathon)

Choose a tag to compare

@tang-vu tang-vu released this 16 Jun 11:34

Keryx is a citation-toll reading agent: ask it a question with a budget, and it autonomously decides which paid content sources to buy over x402, reads enough to answer, writes a grounded answer with citations, and then settles a weighted nanopayment to every source it actually cited — in USDC on Arc.

Built for the Lepton Agents Hackathon (Canteen × Circle, on Arc).

What's in this release

  • The agent brain (lib/agent/) — decompose → discover → decide → fetch → sufficiency → synthesize → attribute → settle, emitting a human-readable rationale for every buy / skip / cache / stop decision.
  • Real x402 settlement on Arc testnet — fetch tolls and weighted citation rewards both settle via Circle Gateway batching (@circle-fin/x402-batching). No mocked settlement in the demo path.
  • Two-tier creator economy — a small access toll on fetch + a citation pool distributed by LLM-assigned contribution weight; multi-author works split the reward automatically.
  • One-click creator onboarding/register → paste an RSS URL → wallet + x402 endpoint created.
  • Live web app at keryx.cc (Next.js 16 · React 19 · Tailwind 4 · shadcn/ui), with a streaming reasoning UI, metrics, and a creator leaderboard.
  • Agent-to-agent mode — Keryx exposed as a paid x402 endpoint.
  • CLI toolingnpm run ask (one query + full reasoning trace), npm run seed (autonomous volume engine), npm run metrics.

Verified traction (real, on-chain)

  • 15 settled payments (11 weighted citations + 4 access tolls) on Arc testnet
  • $0.16 USDC to creators — 100% to creator wallets, 0% platform fee
  • 4 creators earning · 100% reader→payer conversion · avg ~$0.0107 (true nanopayments)

Run it

npm install
npm run ask -- "How do x402 and stablecoins enable autonomous AI agent commerce?" --budget 0.05
npm run dev   # http://localhost:3939

Runs offline with zero keys (heuristic reasoning + simulated payments). Add an LLM key, fund the agent wallet, and set KERYX_FORCE_OFFLINE=0 for real on-chain settlement.

See README.md, DECISIONS.md, and TRACTION.md for details.