# Pink Agentic AI Payments MCP server: let Claude, ChatGPT or Cursor pay vendors and invoices within spending rules

`pink-agentic-payments-mcp` is the MCP server for **Pink Agentic AI Payments** (by PinkWallet). It lets an AI agent in Claude, ChatGPT, Cursor, VS Code, Windsurf or any other MCP client request a payment, such as paying a vendor invoice or topping up API credits, and Pink checks that payment against the company's spending rules (per-agent budgets, single-payment caps, payee allow-lists, human approval above a threshold) **before** any payment credential is issued. The agent never holds a standing card or key.

> **Production pilot:** a private production tenant (decision-only: Pink decides allow, block or ask a person; your own Stripe, card or bank pays) is available for $99/month. In production, payments above an agent's per-payment limit wait for a person's approval. Details and checkout: https://pinkwallet.com/agentic/pilot/


Every payment request gets one of three answers: `allowed` (with a single-use credential), `pending_human` (a person is asked first), or `blocked` (with the reason).

**Sandbox preview: test money only, no signup required to get a key.**

## Can Claude or ChatGPT pay an invoice through MCP?

Yes, with limits you set. Connect Pink as an MCP server, and the agent gets 7 payment tools (listed [below](#tools)). Asked to "pay the Acme invoice for $1,200", the agent calls `pink.check_policy` or `pink.request_payment`; Pink pays only if the invoice fits the rules, asks a person if it needs approval, or refuses and says why. Setup guides: [Claude](https://pinkwallet.com/agentic/connect/claude) · [ChatGPT (developer mode)](https://pinkwallet.com/agentic/connect/chatgpt) · [Cursor](https://pinkwallet.com/agentic/connect/cursor) · [Claude Code](https://pinkwallet.com/agentic/connect/claude-code) · [Walkthrough: an MCP server that lets AI agents pay invoices](https://pinkwallet.com/agentic/learn/mcp-server-for-ai-agents-to-pay-invoices).

## How this package works

This package is a local **stdio** MCP server that bridges any MCP client to Pink's remote Streamable HTTP MCP server (`https://agentic-sandbox.pinkwallet.com/mcp`). It forwards `tools/list` and `tools/call` (and resources/prompts, if the remote adds them) transparently, so the local server always reflects whatever tools Pink's remote server currently exposes; nothing is hardcoded. Clients that support remote MCP can also connect to that URL directly.

## Quick start

### 1. Get a free sandbox agent key

No signup needed. Either run:

```bash
npx -y pink-agentic-payments-mcp --create-sandbox
```

which creates a sandbox workspace and prints a ready-to-paste client config, or create one yourself:

```bash
curl -s -X POST https://agentic-sandbox.pinkwallet.com/v1/sandbox/workspaces
```

and take any `agents[].key` from the response.

### 2. Add it to your MCP client

**Claude Desktop** (`claude_desktop_config.json`), **Cursor**, **Windsurf**, and most other stdio-based MCP clients:

```json
{
  "mcpServers": {
    "pink": {
      "command": "npx",
      "args": ["-y", "pink-agentic-payments-mcp"],
      "env": { "PINK_AGENT_API_KEY": "pk_sandbox_agent_..." }
    }
  }
}
```

**VS Code** (`.vscode/mcp.json`):

```json
{
  "servers": {
    "pink": {
      "command": "npx",
      "args": ["-y", "pink-agentic-payments-mcp"],
      "env": { "PINK_AGENT_API_KEY": "${input:pink_agent_key}" }
    }
  }
}
```

Env vars:

- `PINK_AGENT_API_KEY` (required) — your agent key from a sandbox workspace.
- `PINK_MCP_URL` (optional) — defaults to `https://agentic-sandbox.pinkwallet.com/mcp`.

## Tools

These are live from the remote server's `tools/list` response (see [`TEST-OUTPUT.txt`](./TEST-OUTPUT.txt) for the full run):

| Tool | What it does |
| --- | --- |
| `pink.get_budget` | What this agent may still spend: monthly budget left, spent today, single-payment cap, vault balances, company daily ceiling. |
| `pink.list_payees` | Suppliers, platforms and contractors this company has approved. Paying anyone else stops and asks a person. |
| `pink.list_rules` | The policy rules that apply to this agent, in evaluation order. First match decides: allow, ask, or block. |
| `pink.check_policy` | Dry run. Returns `would_allow` / `would_ask` / `would_block` with the rule and full decision trace. Nothing is spent or held. |
| `pink.request_payment` | The real call. Returns `allowed` with a single-use credential, `pending_human` with a `hold_id` while a person is asked, or `blocked` with the reason. |
| `pink.get_credential` | Poll a pending hold. Once a person approves, this returns the credential. |
| `pink.report_receipt` | File the receipt for a completed payment so it reconciles. |

## How this was tested

`TEST-OUTPUT.txt` in this repo is the full, unedited output of three real runs against a live sandbox workspace (test money only):

1. The built server (`dist/bin.js`) driven over stdio by an MCP SDK test client — `tools/list` plus `pink.request_payment` calls that returned real `allowed` (with an issued sandbox credential), `blocked` (over monthly budget), and `pending_human` (new-payee approval) decisions.
2. The official MCP Inspector CLI (`npx @modelcontextprotocol/inspector --cli`) run headless against the same built server.
3. The same stdio test run again from a packed tarball (`npm pack` then `npx -y ./pink-agentic-payments-mcp-0.1.0.tgz`) in a clean directory, to prove the `npx` install path works end to end.

## Links

- Docs hub: [pinkwallet.com/agentic/developers](https://pinkwallet.com/agentic/developers/)
- MCP connection guide: [pinkwallet.com/agentic/developers/connect-via-mcp](https://pinkwallet.com/agentic/developers/connect-via-mcp/)
- JS/TS client: [pink-agentic-payments](https://github.com/Pink-Agentic-Payments/pink-agentic-payments-js)

## Status

Sandbox preview (test money only). Published on npm as [`pink-agentic-payments-mcp`](https://www.npmjs.com/package/pink-agentic-payments-mcp) — `npx -y pink-agentic-payments-mcp` is all you need, no clone or build step required.

## License

MIT