Pink Agentic AI Payments: sandbox examples

Pink Agentic AI Payments (by PinkWallet, early access) is the approval layer between AI agents and company money: plain-language rules, per-agent budgets and human approvals decide each payment before a one-time card or bank transfer is issued.

Sandbox only: test credentials, no money moves, production not available.

This repo has four tiny, runnable clients against the live public sandbox at https://agentic-sandbox.pinkwallet.com: curl, the official MCP SDK (Node), plain REST (Python stdlib), and a small agent loop that shows all three outcomes an agent can get: allowed, pending_human, blocked.

The sandbox's "coffee shop" template includes a night-time rule (23:00–06:00 local → ask the owner). That means the decision for the exact same payment can change depending on when you run it. Every example below passes local_hour explicitly so the outputs are reproducible, but they still print the decision instead of assuming it, which is the point: policy is evaluated per request, not hardcoded.

Quickstart (2 minutes)

# 1. Create a free sandbox workspace and get an agent key
curl -s -X POST https://agentic-sandbox.pinkwallet.com/v1/sandbox/workspaces \
  -H "Content-Type: application/json" \
  -d '{"company":"My Test Co","email":"","template":"coffee"}'

# 2. Export the agent key it returns (agents[].key, looks like pk_sandbox_agent_...)
export PINK_AGENT_KEY=pk_sandbox_agent_...

# 3. Run any example below
./01-curl/quickstart.sh

Use with Cline

This server works as a remote MCP server (Streamable HTTP). See llms-install.md for the full autonomous setup (create a sandbox workspace, get a key, verify the connection). The Cline config (remote MCP server format) is:

{
  "mcpServers": {
    "pink-agentic-payments-sandbox": {
      "type": "streamableHttp",
      "url": "https://agentic-sandbox.pinkwallet.com/mcp",
      "headers": {
        "Authorization": "Bearer pk_sandbox_agent_REPLACE_ME"
      },
      "disabled": false,
      "autoApprove": []
    }
  }
}

Examples

Example	What it does	Command
01-curl/	Creates a workspace, reads the budget, dry-runs and makes a payment with plain curl + jq	./01-curl/quickstart.sh
02-node-mcp/	Connects over MCP (Streamable HTTP) with the official @modelcontextprotocol/sdk, lists tools, calls pink.check_policy + pink.request_payment	PINK_AGENT_KEY=... node index.mjs
03-python-rest/	Same flow over plain REST using only Python's standard library (urllib); checks, pays, and polls a pending payment	PINK_AGENT_KEY=... python3 pay.py
04-agent-loop/	~60-line loop that buys 3 things: one small order (allowed), one bigger order (pending human approval), one to a blocked category (blocked)	PINK_AGENT_KEY=... node agent-loop.mjs
05-langgraph/	A LangGraph ReAct agent + langchain-mcp-adapters, 3 scenarios (allowed / pending_human / blocked) incl. a prompt-injection probe that gets blocked server-side	PINK_AGENT_KEY=... GEMINI_API_KEY=... python3 05-langgraph/react_agent.py
06-openai-agents-sdk/	An OpenAI Agents SDK agent + MCPServerStreamableHttp, same 3 scenarios, OpenAI-first with a Gemini/LiteLLM fallback	PINK_AGENT_KEY=... OPENAI_API_KEY=... python3 06-openai-agents-sdk/agent.py
07-haystack/	A Haystack Agent + the official mcp-haystack MCPToolset, same 3 scenarios, Gemini chat generator	PINK_AGENT_KEY=... GEMINI_API_KEY=... python3 07-haystack/agent.py
08-race-test/	Fires 40 parallel $7 payments at a single agent's $200/day rule and counts allowed vs. blocked, to check the policy engine holds up under concurrent load	./08-race-test/race_test.sh
09-crewai/	A CrewAI Agent + crewai-tools' MCPServerAdapter, same 3 scenarios, Gemini chat model	PINK_AGENT_KEY=... GEMINI_API_KEY=... python3 09-crewai/crewai_agent.py
10-vercel-ai-sdk/	A Vercel AI SDK generateText tool loop + @ai-sdk/mcp's createMCPClient, same 3 scenarios, Gemini via @ai-sdk/google	PINK_AGENT_KEY=... GEMINI_API_KEY=... npm --prefix 10-vercel-ai-sdk run agent
openapi/	OpenAPI 3.1 spec for the REST API, live-validated against the sandbox; import into Postman or a ChatGPT custom GPT Action	npx -y @redocly/cli lint openapi/openapi.yaml

Each example needs an agent key for the coffee template (a_purch, a_inv, a_mkt, or a_pay: see 01-curl/quickstart.sh to create a workspace and get one).

1. curl: real run, 2026-10-01

$ ./quickstart.sh
== 3. Dry-run a payment (nothing is spent) ==
{
  "decision": "would_allow",
  "reason": "Small supply orders go through",
  ...
}

== 4. Make the real payment (sandbox credential, no money moves) ==
{
  "payment_id": "pay_6c2248fe8a03",
  "decision": "allowed",
  "agent": "Purchasing AI",
  "payee": "Sysco",
  "amount": 150,
  "currency": "USD",
  "rule": "Small supply orders go through",
  "credential": {
    "type": "virtual_card",
    "sandbox": true,
    "max_amount": 150,
    "single_use": true,
    "card": { "pan": "4111 1111 3954 9741", "exp": "12/27", "cvv": "467", "name": "PINK SANDBOX" }
  }
}

2. Node MCP SDK: real run, 2026-10-01

$ PINK_AGENT_KEY=pk_sandbox_agent_hoa8… node index.mjs
== Tools available ==
pink.get_budget, pink.list_payees, pink.list_rules, pink.check_policy, pink.request_payment, pink.get_credential, pink.report_receipt

== pink.request_payment ==
{
  "payment_id": "pay_eb1692583f46",
  "decision": "allowed",
  "agent": "Inventory AI",
  "payee": "Sysco",
  "amount": 150,
  "rule": "Small supply orders go through",
  "credential": { "type": "virtual_card", "sandbox": true, "max_amount": 150, "single_use": true, "card": { "pan": "4111 1111 8247 6657", "exp": "12/27", "cvv": "732", "name": "PINK SANDBOX" } }
}

3. Python REST (stdlib only): real run, 2026-10-01

$ PINK_AGENT_KEY=pk_sandbox_agent_hoa8… python3 pay.py
== Make the payment ==
202 {
  "payment_id": "pay_e6d89cee7384",
  "decision": "pending_human",
  "agent": "Payroll & Bills AI",
  "payee": "Uline",
  "amount": 2200,
  "rule": "Anything over $2,000: ask the owner",
  "hold_id": "pay_e6d89cee7384",
  "approvers_needed": 1,
  "who": "Maya Chen (Owner)",
  "poll": "pink.get_credential(hold_id) · or GET /v1/payments/{id}"
}

== Polling pending payment pay_e6d89cee7384 ==
attempt 1: 200 pending_human
attempt 2: 200 pending_human
attempt 3: 200 pending_human

(It stays pending_human until a person approves it in the console: that's the point of the approval layer.)

4. Agent loop: real run, 2026-10-01

$ PINK_AGENT_KEY=pk_sandbox_agent_hoa8… node agent-loop.mjs
[allowed] p_sysco $80: Small supply orders go through
[pending_human] p_uline $900: Bigger supply orders: store manager checks
[blocked] Quick Gift Cards LLC $50: Never: gift cards, cash-like, crypto

Docs & registry

• Open guide: agentic-ai-payments: protocols, providers, and spending controls, with these examples linked as the framework-integration section
• Developer docs: https://pinkwallet.com/agentic/developers/
• MCP Registry: com.pinkwallet/agentic-payments-sandbox
• Public sandbox: https://agentic-sandbox.pinkwallet.com

Maintained by the PinkWallet team.