v2.3.0: Typed 402s, tokenized assets, MPP on the edge, and two alarms that stopped lying

Thirty PRs merged on 2026-09-02, every one verified on production the same day.

Buyers

• Typed 402s. Every challenge carries the typed output schema as accepts[0].outputSchema (one copy, first accept) as well as in the bazaar extension, and every route validates under @x402/core's own schemas. Stock x402 clients, native MPP buyers and mppx x402 buyers all still settle.
• Tokenized real-world assets. rwa-list, rwa-markets, rwa-asset, rwa-issuers, rwa-issuer: 649 tokenized stocks, ETFs and commodities with onchain market data and the 33 issuers behind them, $0.003 to $0.006.
• Router honesty. executeVia appears only on rows the router will pay now (executeViaCallableNow); every row and index seller carries routerDispatchEligible and routerDispatchReason; a manifest-priced route is read live once and then weekly so a seller's newly added rail reaches the index without a manifest change.
• Models. gpt-4.1-nano and the openai/o4 prefix retired ahead of OpenAI's dates (gpt-5.6-luna is the nano default, gpt-5-nano stays, gpt-5.6-terra joins premium); /v1/images/fast fails over to gpt-5-image-mini. A refused MCP credential answers the spec's -32043.

Sellers

• agent402-tollbooth 0.10.0. The edge build (Cloudflare Workers, Next.js edge, Deno, Bun) takes MPP through the same verifyX402 callback: a WWW-Authenticate: Payment challenge beside the x402 quote, HMAC- and resource-bound credentials translated to PAYMENT-SIGNATURE. The wire codec is a runtime-agnostic module shared with the Node build.
• Index. Solana seller leaderboard scanned incrementally; a learned verb no longer flips a path's other method; row-level network inheritance; Base chain-truth on a refused payment.

Reliability and money-safety

• Tempo subscriptions: a slow-RPC failure retries in minutes, and a send that timed out is settled by reading the chain (memo bound to subscription and period) before anything is signed again.
• In-flight report composites are cut off on SIGTERM, so a deploy never spends upstream and then discards the answer.
• Every skill pack runs every tool it advertises; feed-watch walks the newest feed items instead of betting on the first.
• The Postgres alarm no longer confirms a failure against its own cache; the paid canary proves the Tempo subscription rail after every deploy.
• Security review of the day's changes: one Medium found and fixed the same day. Solana leaderboard evidence (attributed by seller-advertised payTo) no longer counts toward the Base router floor; Solana proven-ness is a pay-time chain read against the accept's own payTo.
• mppx 0.9.2, fast-uri 3.1.7, qs 6.16.0 (four same-day CVEs cleared).

Full detail per change on https://agent402.tools/changelog.

v2.2.0: Agentic Finance, MPP on Tempo, reports for people, and a gateway that bills what the model used

v2.2.0: Agentic Finance, MPP on Tempo, reports for people, and a gateway that bills what the model used

Note on scope: about 140 PRs merged since v2.1.0 (2026-08-17). This is a highlights release, verified against the live site, not an exhaustive changelog.

Catalog today: 559 entries, 485 tools + 74 skill packs (500+ evergreen on all marketing surfaces as usual; this is a release-specific snapshot). The catalog was curated down on measured use: entries with zero external use in 30 days were retired so every remaining tool earns its place.

Positioned under Agentic Finance

Agent402 is the applied layer of Agentic Finance: agents that pay and get paid on their own over open protocols. New pages define and prove it: /agentic-finance (the definition and the stack), /101 (a presenter-mode walkthrough of x402 and MPP with a live demo), /glossary, and /why: seven things that are different about paying here, each linked to the surface that proves it. The same seven points ride in /llms.txt, the MCP instructions, the README and the package docs.

MPP everywhere, with native settlement on Tempo

• Native MPP settlement on Tempo (tempo/charge through Tempo's relay) beside the evm method on Base and Celo; every paid route offers both wires on the same 402. Rejected credentials answer RFC 9457 problem documents.
• Chain-truth confirmation: when a relay reports a broadcast failure for a payment that actually settled, the gate reads the chain and honours it, so a buyer is never told 402 for money that moved.
• MPP subscriptions over tempo/subscription with a gas sponsor; renewals are pulled server-side and proven daily by a live canary.
• Cards over MPP: Stripe's stripe/charge method on routes from $0.50, live-proven.
• The MPP marketplace (/mpp-marketplace): an independent crawler over the mpp.dev registry, MPPScan and our own x402 crawl, verified by real challenges only, ranked by inbound transfers on Tempo read from the chain.
• The Smart Order Router pays MPP sellers on Tempo on the agent's behalf, chain-matched and gated by on-chain proof; agent402-client pays MPP sellers with a stock mppx fetch.
• agent402-tollbooth 0.9.x speaks MPP, settles natively on Tempo with split payments, and carries the same chain-truth confirmation.

Reports, monitors and credits, for people and agents

• /reports: finished, cited deliverables with a data appendix, bought by card or by an agent over x402/MPP: company dossier and ticker pack, insider flow, 13F fund report, filing report, domain security audit, token risk and Solana token brief, deep research (three tiers), market brief, FDA recall report, IPO digest, and a LinkedIn article package with images at LinkedIn's own sizes.
• /monitors ($5 a month): domain, fund, insider, filing, recall, token and IPO watches that probe daily for free and re-run the paid report when the facts change.
• /credits: prepaid credits by card, spent on any paid route with a bearer key, no wallet needed; agent402-mcp 0.13 and agent402-client accept the key.
• Programmatic SEC pages (/reports/insider/:ticker, /reports/fund/:manager, /reports/dossier/:ticker) with real filing teasers, built to be cheap on a paid egress.
• Three rounds of report-input work so every synthesis is handed what the source already has: the XBRL operating-to-net bridge and verbatim filing excerpts, 13F amendments folded in, Schedule 13G holders, derivative Form 4s, recall feeds newest-first, TLS trust in the domain grade, DNS posture and recursive SPF, GoPlus and DexScreener in token risk, page bodies and a citation audit in research. Prices re-derived from measured upstream cost.

The LLM gateway bills what the model used

• Metered tier (POST /v1/metered/chat/completions): every 402 quotes this exact request from its body; exact clients pay the quote, upto and credits buyers settle actual usage x 1.15. Proven daily by the canary's metered-upto leg.
• Three wires on every tier: OpenAI chat, OpenAI Responses and Anthropic Messages; plus rerank, a grounded tier that cites the web on every answer, images at flat per-picture prices, video, and server tools on pro/premium under server-owned cost bounds.
• agent402-openclaw 0.3.1: Agent402 as an OpenClaw model provider, paid by card or wallet, metered by default, upto when the wallet holds a Permit2 allowance; tested against a real OpenClaw install in CI.
• Margin telemetry on streams, provider discounts (prompt cache, flex) passed through at cost, reasoning defaults that avoid paid empty answers, and a live-catalog guard that fails CI on any dead model id.

Reliability and operations

• Stellar facilitator: settlement failover across Soroban RPCs on transport failures, OpenZeppelin as settle fallback, Horizon-confirmed settlement.
• Deploys drain at once (a volume-backed service has nothing to overlap with), boot profiled and cut, the index cache warm-starts incrementally, and CI runs 17 lanes split by measured seconds with a merge helper that trusts only the push-event run.
• Two security review rounds this cycle, each with fixes landed the same day; the retry contract is published in a form a buyer can verify from the response they hold.
• Site redesign: the milled + obsidian system, light by default with a dark toggle, self-hosted Geist, new brand marks, and affirmative copy throughout.

Packages

agent402-mcp 0.13, agent402-client 0.8.1 (adds route(), contributed by @epistemedeus), agent402-tollbooth 0.9.3, agent402-openclaw 0.3.1, agent402-agentkit 0.1.0 (a Coinbase AgentKit action provider). Published packages are verified daily against production from the registry.

Contributors

Thanks to @epistemedeus for seller response contracts, request constructibility and the client route() method.

Payments

USDC over x402 across 11 chains, USDG on Robinhood Chain, native MPP on Tempo, cards over MPP, prepaid card credits, or a few seconds of proof-of-work on the pure-CPU tools. By default (FREE_MODE=true) everything runs free.

Proof

Every paid call settles on a public ledger. Live counters: /api/stats, transactions by rail and wire: /revenue, seller ranking: /api/leaderboard, uptime measured from outside production: /status.

v2.1.0: Redesign, four new chains, hardened site security

v2.1.0: Redesign, four new chains, hardened site security

Note on scope: 273 PRs merged since v2.0.0 (2026-07-14). This release covers the
themes below in detail, verified against the live site. It's a highlights release,
not an exhaustive changelog of every merged PR.

Catalog today: 531 entries, 428 tools + 103 skill packs (500+ evergreen on all
marketing surfaces as usual; this is a release-specific snapshot).

Full site redesign, all 7 primary pages

Homepage, /what-is-x402, /sell, /tools, /leaderboard, /skills, and
/marketplace were all rebuilt on a shared new visual system: restyled nav, a live
world map on the homepage, a real in-browser proof-of-work demo you can run without
a wallet, and a redesigned leaderboard, catalog, and marketplace. A few data-privacy
improvements shipped alongside the redesign, and the leaderboard now excludes
Agent402's own row, consistent with it being a neutral index.

Hardened site security

The site's Content-Security-Policy is now meaningfully stricter, closing off a class
of risk around script execution sitewide. Covered by a new automated test that runs
on every deploy.

Four new payment rails: Celo, Avalanche, Sei, Optimism

Chains accepted for settlement grew from 8 to 12: Base, Solana, Polygon, Arbitrum,
Monad, Celo, Avalanche, Sei, Optimism, Stellar, and Algorand, plus USDG on Robinhood
Chain. Each new rail carries the same reliability bar as the rest: real settlement
proven on-chain and covered by a daily paid canary.

Payments

USDC over x402 across 11 chains, or USDG on Robinhood Chain, or a few seconds of
proof-of-work on the pure-CPU tools, no wallet needed. By default (FREE_MODE=true)
everything runs free.

Proof

Revenue is public: every paid call settles to
agent402.base.eth,
verifiable by anyone on Basescan. Live counters: /api/stats,
seller ranking: /api/leaderboard,
uptime: /status.

v2.0.0 — 500 strong: the curated catalog

v2.0.0 — The 500

Agent402's catalog is now 500 endpoints — 400 tools + 100 skill packs. This is a
major version because roughly 970 routes are gone: if you were calling a pairwise
converter endpoint, read on.

500 strong. Every one tested, priced, and settled on-chain. Every one earns its place.

What changed

Three moves, one number:

1. Pruned ~970 pairwise unit-converter endpoints (/api/convert-miles-to-km and
   friends). They were generated permutations of one unit table. A single parametric
   tool — POST /api/unit-convert — serves every pair with the same unit ids and the
   same math. Catalog: 1,432 → 462.
2. Built 30 new tools + 8 new skill packs — the seven groups below. Every one
   answered its own documented example live (HTTP 200 against the real upstream)
   before it was allowed in. Catalog: 462 → 500.
3. Raised the bar. Every entry answers its own documented example on every CI run,
   is priced to market, and was live-verified before it shipped. The catalog grows only
   when a tool is worth calling.

The quality gate

scripts/sync-count.js --check runs in CI and derives the counts from the booted
server — the total from /health, the pack count from /api/skill-packs.json — and
fails the build if the documented total drifts from reality, or if the catalog falls
below a 400-entry floor (a drop that size means a kit fell off the build, not a
curation decision):

Catalog fell below the 400-entry floor (got N) — a kit is probably missing.

The check can't be gamed by editing a doc (the counts come from the running server),
and it runs before any count sweep, so a broken catalog can never write its own
numbers into the docs.

Graceful retirement

The retired converter routes don't 404. Both shapes the old endpoints had — the slash
form GET /api/convert/<from>-to-<to>?value=N and the slug form
POST /api/convert-<from>-to-<to> — return a teaching 410: the body names the
replacement route and hands back a ready-to-send input ({ value, from, to }, with the
caller's own value and both unit ids parsed out of the old path), plus an
/api/find?q=unit+convert discovery hint. Never a 301 — an agent must not silently
re-POST a paid call across routes; it gets the exact new call instead.

The retired tool pages do 301 to the survivor's page: a page visit has no re-POST
hazard, and inbound links keep their equity.

The quality bar (same bar, smaller surface)

• Every tool answers its own documented example on every CI run — the release
  blocks on any failure.
• The 30 new tools were live-verified against their real upstreams before merge.
• Every tool that touches the network is wallet-only. The free tier is pure-CPU
  only: 216 of the 500 work with no wallet at all, via signed, single-use,
  slug-scoped proof-of-work.
• Deterministic only — no LLM in the serving path. Same input, same output, full
  OpenAPI schemas.

New capabilities — 30 tools + 8 packs, seven groups

Group	What you get
Contract / EVM (7)	contract-source + contract-abi (Sourcify), solidity-scan (fixed heuristic ruleset — a pattern check, not an audit), calldata-decode, selector-lookup, tx-simulate (read-only eth_call, nothing signed), address-label
Market micro-feeds (6)	options-chain, premarket-quote, stock-dividends, dividend-calendar, crypto-orderbook, stablecoin-peg
Entity enrichment (5)	lei-lookup (GLEIF), wikidata-entity, github-repo, gravatar-check, favicon-grab — all keyless
Web / content (4)	feed-parse (RSS/Atom/RDF), unshorten-url (redirect chain, SSRF re-checked per hop), archive-snapshot (Wayback), search-videos
Media / format (5)	image-exif, image-dominant-color, image-crop, json-schema-infer, srt-convert
Locale / time (3)	ics-parse (RFC 5545 + bounded RRULE expansion), public-holidays, country-info (committed dataset, pure CPU, free tier)
Skill packs (8)	contract-audit, tx-forensics, market-open, entity-enrich, feed-watch, schema-guard, subtitle-pipeline, locale-brief — whole-job bundles over the new tools, callable as MCP prompts

Payments (unchanged)

USDC over x402 on Base, Solana, Polygon, Arbitrum, Monad, Stellar & Algorand, plus
USDG on Robinhood Chain — 8 chains. Or a few seconds of proof-of-work on the pure-CPU
tools, no wallet needed. By default (FREE_MODE=true) everything runs free.

Proof

Revenue is public: every paid call settles to
agent402.base.eth,
verifiable by anyone on Basescan. On-chain settlement receipts for the new tools land
with the launch sweep — a real paid buy against each new endpoint, so every entry in
the table above is advertised only after it has actually settled. Live counters:
/api/stats · seller ranking:
/api/leaderboard.

v1.3.0 - 1,431 tools: federal-data pack, market-pulse, reliability invariants

1,431 deterministic, pay-per-call web tools an AI agent can call and pay for per request - USDC via x402 across multiple chains, or free via proof-of-work.

New tools (+24 since v1.2.0)

Federal open-data pack (11 tools, all keyless and public-domain):

• drug-recalls, food-recalls, device-recalls, drug-adverse-events - FDA recall/enforcement + FAERS adverse events (openFDA)
• vin-decode, vehicle-recalls - NHTSA vPIC decode + safety recalls
• college-lookup - US College Scorecard (tuition, admit rate, size)
• fec-candidates - US federal election candidates (FEC)
• federal-awards - US federal contract awards (USAspending)
• geo-lookup - lat/lon to county/state/census block (FCC)
• fema-disasters - FEMA disaster declarations by state (openFEMA)

x402-market-pulse - cross-provider x402 market sentiment: top sellers ecosystem-wide by real on-chain USDC settled (multi-lens: revenue / distinct buyers / calls), plus the supply mix classified into a functional taxonomy.

Reliability

• Semantic self-check invariants - the production self-check now verifies monitored tools return the correct answer (a fixed VIN decodes to the right car, a historical recall never un-happens, national debt only grows), not merely "no error." tool-alert pages on wrong-answer regressions, every 30 min.

Chains and discovery

• Monad added as an x402 rail (USDC on Monad via the molandak facilitator), proven with a real on-chain settlement.
• PayAI added as a third seller-discovery source for the /index crawler (non-Base breadth, Solana-heavy), with testnet + placeholder-origin filtering.

Platform

• Server-side pageview analytics - privacy-preserving: a pseudonymous, daily-rotating, salted one-way visitor key. No client JS, no cookies, no CSP change, no raw IP/UA stored, no cross-day tracking.
• MCP connector polish - consistent snake_case tool names + Returns { ... } output-shape hints on both the hosted /mcp connector and the agent402-mcp npm package (CallTool still accepts the raw slug, so nothing breaks).
• safeFetch gained SSRF-safe POST support (the SSRF guard is method-agnostic).
• Performance - chain/marketplace pages hoist repeated inline styles to CSS classes (decompressed HTML down ~54% on the busiest chain; no visual change).

v1.2.0 — 1,407 tools + 100 skill packs

What's new since v1.1.0

100 skill packs (up from 51)

• 49 new packs across finance, crypto, security, and dev tools — each a whole agent job in one x402 payment, also callable as MCP prompts.
• Full quality pass: every pack composes ≥3 real deterministic tools with a verified multi-step workflow (no stubs).

1,407 tools (up from 1,358)

• company-financials, multi-search, and more added to the deterministic, flat-priced catalog. No LLM in the serving path.

Neutral x402 Index

• Committed operator seeds so external x402 sellers that can't reach the CDP Bazaar can still be listed and routed via /index + /api/route.
• Health-aware Smart Order Router ranks the cheapest healthy tool across every crawled seller.

Live revenue transparency

• /revenue is linked from the site — every external payment is verifiable on-chain, each linked to its explorer proof; the page leads with the verifiable payment count.
• Daily revenue digest hardened for reliable reads across all 6 rails.

Numbers

• 1,407 tools + 100 skill packs
• 6 payment chains — Base, Solana, Polygon, Arbitrum, Stellar (USDC) + Robinhood Chain (USDG)
• 1,181 tools free via proof-of-work (no wallet)
• External revenue verifiable live at https://agent402.tools/revenue

v1.1.0 — 6 chains, premium packs, Stellar, Stripe ACP

What's new

6 payment chains

• Stellar added as 6th chain (OpenZeppelin facilitator, fees sponsored)
• All 6 chains proven with real on-chain settlements: Base, Solana, Polygon, Arbitrum, Stellar, Robinhood Chain
• Revenue page now shows transactions + timestamps on all 6 chains

Premium skill packs ($0.25–$0.50)

• company-dossier ($0.50) — quote + financials + SEC filings + insider trades + news
• domain-intel ($0.25) — WHOIS + DNS + TLS + headers + tech stack + robots.txt
• crypto-dossier ($0.30) — price + 90d history + trending + global + news + extract
• financial-analysis ($0.08) and market-brief ($0.05) — lighter bundles

New tools

• company-financials ($0.02) — 9 key metrics from SEC XBRL in one call
• multi-search ($0.08) — 2–5 web searches in parallel

Stripe ACP endpoint

• /acp/feed + /acp/manifest — full 1,358-tool catalog as Agentic Commerce Protocol product feed

Infrastructure

• PostHog payer wallet tracking on tool_call + payment_settled events
• Nasdaq relay worker deployed (fixes earnings-calendar timeouts)
• Whois RDAP retry on timeout
• Cloudflare integration docs, starter templates (LangChain, CrewAI, Claude Code)
• Integration playbook for Circle, x402 Foundation, AWS AgentCore

Numbers

• 1,358 tools + 51 skill packs
• 6 payment chains (Base, Solana, Polygon, Arbitrum, Stellar, Robinhood)
• 63/63 canary verified on prod
• $7.19 external revenue from 1,517 payments all-time

v1.0.0 — 1,338 tools + 42 skill packs

Agent402 v1.0.0 — the open x402 index + 1,338 pay-per-call tools + 42 skill packs for AI agents.

Highlights:

• 1,338 deterministic web tools (browser, search, OCR, PDFs, finance, crypto, EDGAR, stats, memory, 1000+ utilities)
• 42 curated multi-tool skill packs as MCP prompts
• x402 Index: Find, Route (Smart Order Router), Leaderboard (on-chain ranking)
• Base Builder Code verified — onchain attribution via ERC-8021
• 8 framework adapters: OpenAI, Anthropic, Vercel AI SDK, LangChain, LlamaIndex, Strands, Google ADK, OpenAI Agents
• agent402-tollbooth: open-source pay-per-crawl gate
• agent402-client: buyer SDK with auto-payment and builder code registration
• Base Notifications API integration
• Base MCP plugin spec for native Base MCP agent access
• Free via proof-of-work or pay per call in USDC on Base