# LLM Wallet Router

> Proof of concept: choose a model, prepare an answer, pay its actual input/output token invoice with x402, receive the answer. No router account, API key, subscription or Gateway deposit.

## Endpoints

- Origin: https://x402-direct.46-225-27-102.sslip.io
- MCP (Streamable HTTP): https://x402-direct.46-225-27-102.sslip.io/mcp
- Models and tariffs (free): https://x402-direct.46-225-27-102.sslip.io/v1/models
- Chat completions: POST https://x402-direct.46-225-27-102.sslip.io/v1/chat/completions
- Real proof, when published: https://x402-direct.46-225-27-102.sslip.io/proof.json

## Payment flow

1. Discover the model and unit prices. The quote tool is a free estimate, not a payable invoice.
2. Call chat without payment. The server prepares and stores the bounded answer privately, then returns an exact x402 v2 invoice for its actual provider-reported input/output usage.
3. A compatible Base USDC wallet checks the invoice against its budget and signs the EIP-3009 exact amount. Preserve the private invoice URL in the payment payload.
4. Retry the same request with the payment. The server verifies/settles once, then releases the answer and receipt. Failed generation is not charged. A pending/unknown payment must be reconciled, never blindly paid again.

HTTP uses PAYMENT-REQUIRED / PAYMENT-SIGNATURE / PAYMENT-RESPONSE. MCP carries the payment in tools/call params._meta["x402/payment"] and the receipt in result._meta["x402/payment-response"]. Ordinary MCP registration does not sign payments automatically.

Preserve a private random UUIDv4 request_id (MCP) or Idempotency-Key (HTTP) across retries. Invoice URLs and recovery keys are private capabilities: do not publish them. Paid recovery must not cause another charge.

## PoC tariff and limits

Confirm current values in /v1/models: beta tariff is 1 USDC per million input tokens, 3 USDC per million output tokens, plus 0.002 USDC per completed request. Maximum client budget: 0.01 USDC/request. This is an introductory selling tariff, not a claim of provider pass-through cost.

Text messages only, bounded input/output, no full OpenAI feature-parity claim. EIP-3009 exact payments need no token allowance or Gateway top-up. The server bears unpaid generation work in this PoC and caps concurrent/outstanding work.

This is a technical proof of concept, not a production availability guarantee. Only published transaction-backed results count as successful paid tests.