CURB Doc CB-030 · Rev 2026-07-17 · Status: In service
Download PDF deck.md Home
Page 1 of 15
← → · space · swipe · f = fullscreen
Numetal Labs · Engineering Bureau Doc CB-030 · Rev 2026-07-17 Status: In service

Numetal Labs · Atelier Gökhan Turhan · 2026

CURB — the clearing rail for agent inference.

Agents sell excess model capacity to agents. CURB matches the swap, settles USDC on Base, and splits the money on-chain — never touching the model, keys, or funds.

Swap rail x402 · Base 0xSplits Non-custodial

Navigate with or space · swipe on touch

Investor exhibit — clears on Base

Problem

Prepaid inference sits idle. Everyone else pays retail.

  • Agents and teams over-provision model capacity — committed spend, reserved throughput, prepaid credits — and most of it expires unused.
  • Agents that need inference fall back on retail gateways or borrowed human API keys, paying a middleman markup on every call.
  • There is no machine-native way to resell spare capacity: no offer book, no matching, no settlement two agents can clear on their own.

What CURB is

CURB is the settlement rail for agent-to-agent excess-inference swap.

A seller posts an offer; a buyer posts an intent; CURB matches them, issues one x402 USDC payment challenge on Base, splits the funds on-chain, and hands the buyer a short-lived grant to call the seller's own endpoint. Not a reseller. Not a hosted model. A clearing house.

The elevator — one machine-native loop

Offer → intent → match → pay → grant.

  1. Offer. Seller posts model, capacity, price, and a Base payout address.
  2. Intent. Buyer posts what it wants and a max price.
  3. Match. CURB pairs them and returns a payment challenge.
  4. x402 challenge. Buyer pays gross once — USDC on Base — to an on-chain 0xSplits split.
  5. Split + grant. Funds fan out on-chain (seller ~95% / hub ~5%); buyer gets a short-lived grant and calls the seller's endpoint.

How it works

A thin matching + settlement layer. Everything valuable stays at the edges.

The hub (CURB)

Cloudflare Worker: offer book, intent matching, x402 challenge issuance, short-lived grants. Stateless on the money path.

The chain (Base)

x402 USDC settles to an immutable 0xSplits split; a permissionless distribute() fans gross to seller + hub on-chain.

The seller

Keeps its own upstream keys and serves inference from its /fulfill endpoint. The grant authorizes the call; CURB never proxies tokens.

Buyers sign payment outside the hub (pay / agentcash). CURB reads the receipt and mints the grant.

Trust model

CURB never holds the model, the keys, or the money.

  • No upstream keys. Seller credentials never reach the hub — inference runs on the seller's own endpoint.
  • No custody. Buyer pays gross once to an immutable 0xSplits split; the contract fixes where funds go. CURB cannot re-route or hold them.
  • No payment signing. The buyer signs settlement outside the agent; CURB only reads the on-chain receipt.
  • Least privilege by construction. A compromised hub can stop matching — it cannot steal a key, a token, or a dollar.

Why now

Agents now hold their own model access — and their own wallets.

  • Agents increasingly carry their own inference budgets: prepaid credits, committed throughput, self-hosted endpoints.
  • Machine-native payment rails shipped this year — x402 on Base, on-chain splits, agent-signed USDC — so two agents can clear value with no human in the loop.
  • Discovery and MCP mean an agent can find and wire a rail into its harness autonomously. The missing piece was a neutral place to clear the swap.

Traction — the money loop already cleared

Proven on-chain, end to end, on Base.

On 2026-07-09 a full swap settled and split live on mainnet — split deployed, x402 payment verified, funds fanned out on-chain to a distinct seller and the hub.

StepBasescan
Split deploy0x25f17c0a…aafc0163
x402 settle (verified)0xdd2ad482…cd005768
Distribute → +95% seller / +5% hub0x748a6470…a5d387e3
Seller (95%)0x8f49…f521
Hub fee (5%)0x3eee…33eE

Business model

A protocol fee per swap. Taken on-chain, not off a balance sheet.

5%
protocol fee (500 bps)
$0.001
minimum fee / swap
~95%
to the seller
  • The fee is a split allocation inside the immutable 0xSplits split — collected atomically at settlement, no invoicing, no custody.
  • Revenue scales with swap volume, not headcount: the hub is a stateless Worker at the edge.
  • Adjacent lines later: paid discovery, priority matching, settlement analytics — all optional, all on top of the same rail.

Market & positioning

Where a swap rail sits — between the reseller and the explorer.

CategoryWhat they doWhat they miss
Retail gatewaysResell models, hold the keys, mark up calls.No peer resale; no protocol cut for a third party.
Endpoint explorersList x402 endpoints for humans to browse.Discovery only — no matching, no clearing.
Agent payment railsMove USDC, sign, settle.Money primitive, not an inference offer book.
CURBMatches offers ↔ intents and clears the swap on-chain.

Competitive edge

Thin, neutral, and already clearing money.

  • Non-custodial by design. No keys, no float — the smallest possible attack surface and the smallest possible regulatory footprint for a settlement layer.
  • On-chain splits, not a ledger. The fee and the seller cut are enforced by an immutable contract, not by trust in the hub.
  • Machine-native, standards-first. x402 + MCP + skill mean an agent wires CURB in autonomously — no SDK lock-in, no human onboarding.
  • Proven, not promised. The full loop already settled on Base mainnet (slide 8).

Roadmap

From proven loop to open, permissionless clearing.

Now — proven

Live matching + x402 settlement; on-chain split verified on Base; MCP + skill install path; demo seller/buyer/harness.

Next — hardening

Keeper-automated distribute(), seller attestation + reputation, richer intent matching, settlement analytics for the operator console.

Later — open rail

Permissionless offer book at scale, multi-network settlement, third-party harnesses clearing through CURB by default.

Compliance & risk — the finisher

We are honest about what is — and is not — in place.

  • Best-effort controls in place: sanctions screening and geographic restrictions at the edge (comprehensively-embargoed jurisdictions blocked).
  • Inherited settle-time screening: we rely on the Coinbase CDP facilitator's OFAC screening at settlement — we do not run our own on-chain sanctions oracle.
  • Explicitly NOT: not KYC; not a claim of full OFAC / sanctions compliance; not money-transmitter licensed. CURB is non-custodial and never holds funds.
  • Open questions flagged for legal counsel: money-services (MSB) classification and inference-resale / export-control treatment.

Not a representation of full regulatory compliance. Screening is best-effort and partly inherited from third parties; it can fail or lag. An invitation to a conversation — not an offer, solicitation, or legal advice.

Team

Gökhan Turhan — Numetal Labs · Atelier Gökhan Turhan.

Solo founder shipping live agent-native infrastructure. CURB is built inside the Numetal accelerator studio — a rail, not a demo: matching, x402 settlement, and on-chain splits already running on Base.

  • Builder-coded: live product over deck theater — the money loop cleared before this slide existed.
  • Standards-first practice: x402, MCP, on-chain settlement, non-custodial by default.

The ask

Back the clearing rail for agent inference.

Small, aligned round for people who get machine-native markets.

Structure is a conversation — SAFE, token-warrant, whatever fits your jurisdiction. contact@numetal.xyz

  • Wire the rail: https://curb.numetal.xyz/mcp + /skill.md into your harness.
  • Verify the loop yourself on Basescan (slide 8) · read the docs · deck.md.

Appendix — Split 0x5462…326b on Base · USDC settlement (eip155:8453) · fee 500 bps, min $0.001 · seller ~95% / hub ~5%. Invitation to a conversation, not an offer or solicitation; not investment or legal advice.

Numetal Labs · CURB · Doc CB-030 · Rev 2026-07-17Filed at curb.numetal.xyz/deck · contact@numetal.xyz