# agents.u.cash — Roadmap

A phased plan that proves the loop with the smallest valuable slice, then compounds leverage, then aims to own the economy — **non-custodial from day one**. Every phase preserves the invariant: the platform never holds user funds or user keys, so it is never a money-service business.

## v1 — prove the loop (weeks)

Ship the smallest end-to-end slice: an agent signs up, prices a resource, and gets paid over 402 in crypto — settled **directly to the agent's own wallet** — while a human can pay the same resource by card through the merchant's own processor.

- **Agent self-serve signup** on `agents.u.cash`: wallet-first, **$0 start** (no payment to begin; the tenant DB is provisioned lazily on first write) + per-IP/wallet throttling + an OFAC sanctioned-address blocklist. Verifying an email is an **optional reward** that grants free starter credit (`CLOUD_INITIAL_FREE_CREDIT_AGENTS`).
- **A dedicated, cleanly-branded agent REST API** with opaque identifiers — agents never see internal ids or the legacy RPC surface.
- **The multi-rail 402 challenge:** a `detect` entry (every address-based coin the agent has a wallet for — reuses the engine's detection, zero new crypto) + a `redirect` entry for fiat/card routed to the **agent's own** Stripe processor (self-service: agents connect their own Stripe account via `POST /v1/stripe`, which liveness-checks the key + product before storing) — both paid **directly buyer→seller**. An `exact` EVM entry (x402-style signed-payload push, EIP-3009 `transferWithAuthorization`) is **shipped for USDC on Base**: the platform verifies the buyer's signed authorization and settles it **gasless by default** — the platform relayer submits the buyer-signed authorization on-chain (on by default, opt-out), billing the seller gas + 21% per settle; a seller can opt out (`POST /v1/exact-relay {enabled:0}`) to verify-only (buyer-pays-gas). The relayer is non-custodial: it signs only a gas-paying wrapper around a buyer-authored message it cannot alter, never the value move. `detect` still covers every other address-based asset without a relayer, a client-supplied proof, or a funded Base key.
- **Non-custodial verify from launch** — the platform detects the buyer's real on-chain payment to the seller's own wallet and confirms it; it holds no funds and no user keys (no relayer, no client proof). Mainnet crypto is fine at v1 *because* settlement is non-custodial (no custodial sink, no Coinbase account).
- **Fee from the seller's prepaid, non-withdrawable credit** at the platform's UCASH fee rate — the buyer pays the full gross amount direct to the seller; the fee is debited from the seller's credit. No held balance.
- **Node-generated addresses + auto-sweep disabled by default**, gated behind an explicit "this makes you a money transmitter" warning in admin. The agent surface never uses them; agent receive addresses are the agent's own wallets.
- **Machine-friendly docs:** an `openapi.json`, an `llms.txt`, and a three-step quickstart.
- **Agentic discovery (UCP):** each merchant publishes a machine-readable catalog (`/catalog.json`, schema.org/Product) + a [UCP](https://ucp.dev) profile (`/.well-known/ucp`, the open Universal Commerce Protocol standard, not Stripe) so AI shopping agents can find + buy their items. The full REST surface is live: checkout sessions (`POST/GET /checkout-sessions`, `complete`, `cancel`) with **multi-item carts** (**mixed-currency** capable: per-item original currency + a converted cart total; a session reaches `completed` only once every line item settles, and orders report per-item fulfillment), catalog REST (`/catalog/search`, `/catalog/lookup`, `/catalog/product`), and orders (`GET /orders/{id}`), all settling non-custodially on the existing `detect` + `redirect` rails. Responses are **RFC 9421-signed** (ES256) with the merchant key published in the profile, and signed inbound requests are passively verified. **AP2 mandates** (`dev.ucp.shopping.ap2_mandate`) are supported merchant-side: every checkout response carries a detached-JWS `merchant_authorization` (over the RFC 8785-JCS-canonicalized checkout) + a per-session `nonce`, and on `complete` a buyer's `checkout_mandate` (SD-JWT-VC) is verified for issuer signature + checkout-hash binding + **key-binding JWT holder proof** (RFC 9901: cnf-key signature, nonce, aud, sd_hash, iat) + **open-mandate constraints** (max_amount / nbf / allowed_merchant; `mandate_scope_mismatch` on violation), else rejected 401. Remaining (future): identity-linking, the RFC 9421 enforcement toggle, per-attempt fresh nonces + a spec-defined AP2 constraint schema.

**Use cases unlocked:** pay-per-call APIs, pay-per-inference, monetized MCP skills, search/RAG, and agent-as-buyer over crypto.

## v2 — compound leverage (months)

The phase where the platform becomes the settlement layer for an economy, not just a per-call rail — still fully non-custodial.

- **Verify-now / relay-later of signed intents** (the `exact` rail now exists): accumulate *signed messages* (not funds) and batch-relay them, driving effective per-call cost toward sub-millicent for high-volume agents. (No pooled balance, ever.) Today's `detect` rail settles each on-chain payment independently.
- **Discovery:** a Bazaar-grade index of endpoints, pricing, capabilities, and reputation, plus a per-store `/v1/manifest`.
- **Builder Code (ERC-8021) atomic revenue-share** — one inbound payment splits to provider + reseller + platform in a single settle; every agent can be a distribution channel.
- **Multi-chain `accepts`** beyond Base — Solana, Polygon, Tron — with CCTP cross-chain settlement (direct buyer→seller, no platform bridge).
- **Payment-Identifier and Batch extensions** for idempotent, high-volume flows.
- ~~An MCP server and SDKs for the major agent frameworks~~ **Shipped**: the MCP server (79 tools) plus zero-dependency JS and Python SDKs.
- **Agent-operated storefronts** — buy wholesale over 402, sell retail via the merchant's own card processor, each side non-custodial.
- **Smart-contract escrow** (the platform never controls escrowed funds) for pay-per-outcome and trustless deals.
- **Operational hardening:** Stripe dispute handling, PayPal signed webhooks, async webhook fan-out.

## v3 — own the economy (6–12 months)

- **Reputation and trust graph** — every settled 402 is a graph edge; reputation queries inform counterparty choice. Pair with emerging authorization standards (AP2 / Verifiable Intent) at the protocol boundary.
- **Agents paying humans** — the agent pays from its **own** wallet to a human's wallet/email the platform merely relays as information (the platform never signs the payout).
- **Outcome-based pricing and auctions.**
- **A streaming rail** (Solana-native) for continuous services.
- **Consortium and treasury agents** (shared multi-sig, pro-rata distribution — all on-chain, non-custodial).
- **The DB sharding router** — the per-tenant database model must scale before roughly 50,000 active agents; a multi-quarter effort that must start in v3 at the latest.

## Decision dependencies

The roadmap assumes a few product decisions, all now settled (see [RISKS.md](RISKS.md)):

1. **Non-custodial everywhere** — the platform detects/verifies; funds move direct buyer→seller; no held balances, no user keys. This is what keeps the platform off the MSB ledge at every phase.
2. **Fee = the UCASH fee rate, debited from the seller's prepaid non-withdrawable credit** (no float/yield, no held balance).
3. **Anti-spam = $0 start + lazy provisioning + throttle + sanctions** (email-verify is an optional reward for free credit, not a gate).
4. **Node-generated addresses + auto-sweep disabled by default + MSB warning.**
5. **Settlement = on-chain detection** (the `detect` rail) + gasless relay (the `exact` rail, default-on/opt-out). The `exact` relayer is non-custodial: it signs only a gas-paying wrapper around a buyer-authored message it cannot alter, never the value move itself.

## Read next

- [ARCHITECTURE.md](ARCHITECTURE.md): the protocol (detect + redirect model, and the `exact` EIP-3009/x402 rail).
- [RISKS.md](RISKS.md) — the non-custodial legal argument and the residual state-level exposure.
- [VISION.md](VISION.md) — the economy this roadmap builds toward.