Payment Processing Skill

Governed payment operations for AI agents — intent-based payments with approval gates, evidence collection, reconciliation, and full audit trails. Agents propose, policy decides, humans approve, system executes.

What This Skill Does

This skill ensures AI agents never execute payments directly. Every financial action follows the intent → evidence → approval → execution pipeline. No shortcuts, no exceptions.

Workflow	Tool Calls	What It Achieves
Checkout (collect money)	3-4	Intent + evidence + approval gate
Refund	4-5	Verify original + intent + evidence + approval
Payout (pay vendor)	3	Intent + evidence + mandatory approval
Status Check	1	Current state with reconciliation info
Reconciliation	2-3	Match payments to invoices/orders

The Governance Model

Agent creates intent (Draft)
    ↓
Agent attaches evidence (invoice, order, contract)
    ↓
Agent requests approval (Draft → PendingApproval)
    ↓
Human reviews and approves/rejects
    ↓
Agent executes ONLY if Approved
    ↓
Agent reconciles against source record

Key safety features:

Idempotency keys prevent duplicate charges
All amounts in minor units (no floating point errors)
Evidence required before approval can be requested
Payouts ALWAYS require human approval
Full audit trail on every action

Installation

# Claude Code
git clone https://github.com/zavora-ai/skill-payment-processing.git \
  ~/.skills/skills/payment-processing

# ADK-Rust
cp -r payment-processing /path/to/project/.skills/skills/

Requirements

Required:

mcp-payments server connected (v2.0.0+)

Optional:

mcp-finance — invoice creation and payment recording
mcp-notifications — approval alerts to finance managers
mcp-slack — team visibility on payments
mcp-email — customer receipts and refund confirmations

Folder Structure

payment-processing/
├── SKILL.md                       # Main skill (loaded on trigger)
├── scripts/
│   └── validate_payment.py        # Amount conversion + validation + approval prediction
├── assets/
│   └── payment-status-report.md   # Output template for status checks
├── references/
│   ├── tool-sequences.md          # State machine + exact tool patterns
│   ├── cross-mcp-workflows.md     # Payments + Finance + Notifications + Email
│   └── examples.md                # 5 scenarios including duplicate prevention
├── README.md
└── LICENSE

Example

User: "Charge Acme Corp $500 for their Pro plan"

Agent behavior:

Creates checkout intent (50000 minor units, idempotency key generated)
Attaches order reference as evidence
Submits for approval (amount > $100 threshold)
Reports status to user

Result:

✅ Payment intent created and submitted for approval
Intent: pi_001 | Amount: $500.00 | Status: ⏳ Pending Approval
The payment requires human approval (amount > $100 threshold).

Success Criteria

Metric	Target
Governance compliance	100% of payments pass through approval gates
Duplicate prevention	0 duplicate payments (idempotency enforced)
Evidence attachment	100% of intents have evidence before approval
Trigger rate	90% on payment-related queries

MCP Server Compatibility

Designed for mcp-payments v2.0.0+:

Feature	Supported
Intent-based model	✅
Approval gates	✅
Idempotency enforcement	✅
Evidence attachment	✅
Reconciliation	✅
Minor-unit amounts (i64)	✅

Contributors

James Karanja Maina

License

Apache-2.0

Part of the ADK-Rust Enterprise skills ecosystem.

Built with ❤️ by Zavora AI

Skill Payment Processing

README