AgentPay VN

AgentPay VN

VietQR payment infrastructure for AI agents — collect money inside any conversation.

AgentPay VN lets AI agents (Claude, GPT, custom bots) generate payment QR codes, send them to users, and automatically confirm when the money arrives — all without ever holding or touching funds. Money flows directly from the payer's bank account into the merchant's account; AgentPay only reads the bank transaction feed to confirm settlement.

Status: Early access / self-hosted — running on the same swarm as Sổ Nợ AI.

How it works

AI Agent                   AgentPay API              Bank feed (SePay)

1. Create — agent calls POST /v1/payment-requests → gets a VietQR image URL and a checkout page.
2. Send — agent embeds the QR image or sends the checkout link to the user in chat.
3. Await — agent calls await_settlement() (or the MCP tool) to poll until status = settled.
4. Deliver — only after confirmed settlement does the agent release the goods/service.

AgentPay never holds money. The QR points directly at the merchant's bank account number. The platform only monitors the bank transaction feed to detect matching transfers.

Quick start

1. Install

pip install agentpay-vn

2. Get an API key — no account needed

A sandbox key is self-serve and instant:

curl -X POST https://agentpay.servicesai.vn/v1/sandbox/register \

-H "Content-Type: application/json" -d '{"shop_name": "My Bot"}'

-> { "api_key": "ap_test_...", "expires_at": "..." }

export AGENTPAY_API_KEY=ap_test_xxx

From Python: register_sandbox(). From an agent: just call the register_sandbox MCP tool — it issues the key and applies it to the session.

Sandbox keys last 7 days, move no real money, and renew by calling the same endpoint with the same email. For real payments you need an ap_live_* key, which requires a registered merchant with a Vietnamese bank account — sign up at agentpay.servicesai.vn.

3. Collect a payment (3 lines)

from agentpay.client import AsyncAgentPayClient, await_settlement

import asyncio

async def main(): async with AsyncAgentPayClient("ap_test_xxx") as client: pr = await client.create_payment_request(amount=50_000, description="Order #1") print(pr["checkout_url"]) # send this link to your user result = await await_settlement(client, pr["id"], timeout=120) assert result["status"] == "settled"

asyncio.run(main())

See examples/quickstart.py for the full runnable version.

MCP server setup

AgentPay ships an MCP server so any MCP-compatible AI agent can call it as a tool — no extra code needed.

Claude Desktop / Claude Code

Add to claude_desktop_config.json (or use examples/claude_desktop_config.json):

{

"mcpServers": { "agentpay": { "command": "python", "args": ["-m", "agentpay.mcp_server"], "env": { "AGENTPAY_API_KEY": "ap_test_xxx", "AGENTPAY_BASE_URL": "https://agentpay.servicesai.vn/v1" } } } }

Or use the installed console script:

{

"mcpServers": { "agentpay": { "command": "agentpay-mcp", "env": { "AGENTPAY_API_KEY": "ap_live_xxx" } } } }

Available MCP tools

Agent skill (Claude Code, Cursor, and other agents)

The MCP server gives an agent the *tools*. The bundled skill gives it the *judgement*: when to charge, how to format a VND amount, and the rule that nothing ships before status = settled. See skills/agentpay-vn/SKILL.md.

This repository is also a Claude Code plugin marketplace, so the skill and the MCP server install together in one step:

claude plugin marketplace add phuocdu/agentpay-vn

claude plugin install agentpay-vn@agentpay-vn

Claude Code prompts for your API key (stored in the OS credential store, not in settings.json). Leave it blank to start in sandbox — the agent will call register_sandbox on its first charge.

For other agents, copy skills/agentpay-vn/ into wherever that agent reads skills from, and add the MCP server as shown above.

Python SDK

Synchronous

from agentpay.client import AgentPayClient

with AgentPayClient("ap_live_xxx") as client:

Create

pr = client.create_payment_request( amount=150_000, description="Consulting session 30 min", ttl_minutes=30, idempotency_key="session-abc-123", )

Poll manually

import time for _ in range(60): pr = client.get_payment_request(pr["id"]) if pr["status"] != "pending": break time.sleep(5)

Reconcile

txns = client.list_transactions(limit=10)

Asynchronous

from agentpay.client import AsyncAgentPayClient, await_settlement

async with AsyncAgentPayClient("ap_live_xxx") as client: pr = await client.create_payment_request(amount=75_000, description="eBook download") result = await await_settlement(client, pr["id"], timeout=300) if result["status"] == "settled": send_download_link(result["metadata"].get("email"))

Webhook verification

import hashlib, hmac

def verify_webhook(raw_body: bytes, signature_header: str, secret: str) -> bool: expected = hmac.new(secret.encode(), raw_body, hashlib.sha256).hexdigest() return hmac.compare_digest(expected, signature_header)

Register a webhook endpoint:

ep = client.register_webhook(

url="https://your-server.com/webhooks/agentpay", events=["payment.settled", "payment.expired"], ) print(ep["secret"]) # store this — shown only once

API reference

• OpenAPI spec: agentpay-openapi.yaml
• Base URL: https://agentpay.servicesai.vn/v1
• Authentication: Authorization: Bearer ap_live_xxx (or ap_test_xxx for sandbox)

Key endpoints

Self-hosting

AgentPay runs as part of the Sổ Nợ AI FastAPI backend.

Requirements

• Docker Swarm cluster (same as Sono)
• MongoDB (shared with Sono)
• SePay bank feed account (for live payments)
• Nginx with an agentpay.servicesai.vn vhost

Environment variables

Create an API key (admin)

curl -X POST https://sono.servicesai.vn/api/admin/agentpay/keys \

-H "Authorization: Bearer " \ -H "Content-Type: application/json" \ -d '{"org_id": " ", "name": "My bot", "livemode": true}'

The response includes the full key — store it immediately; it is shown only once.

Rate limits

Design principles

1. No money held — QR codes point directly at the merchant's bank account. AgentPay only reads the transaction feed; it never touches the money.

1. Idempotency — pass an Idempotency-Key header on POST /payment-requests to safely retry without creating duplicates (24-hour deduplication window).

1. HMAC webhook verification — every outbound webhook is signed with HMAC-SHA256(whsec_..., raw_body) in the AgentPay-Signature header. Always verify before processing.

1. Sandbox — use ap_test_* keys and POST /v1/sandbox/simulate-settlement to develop and test without real transactions.

1. Minimal trust surface — the MCP server is a thin REST client with no local secrets beyond the API key. Compromising an agent key only exposes one tenant's payment-request creation ability.

License

MIT © 2026 ServicesAI — see LICENSE.