PayOps Agent Mesh — Security skill for Claude Code

PayOps Agent Mesh

Governed agentic workflows for payments operations — reconciliation, audit prep, and financial analytics — built as a direct, hands-on match to an AI Automation Engineer role scoped around an AI-native Payments Operating System: Claude/Anthropic APIs, Python, n8n, and MCP, with risk-tiered approval gating and a tamper-evident audit log.

Runs entirely offline out of the box (no Anthropic API key required to clone, test, or demo it) and switches to real Claude narratives the moment a key is supplied — see [Design decisions](docs/architecture.md#design-decisions-worth-calling-out).

What it does

Endpoint	Capability
POST /reconcile	Matches bank vs. internal-ledger transactions by reference; classifies exceptions (bank-only / ledger-only / amount mismatch); queues amount mismatches for human approval before any correction
POST /audit-report	Turns a reconciliation summary into an audit-ready narrative
POST /analytics-summary	Transaction volume + z-score anomaly detection with a narrative write-up
GET/POST /approvals/*	Human-in-the-loop approval queue for high-risk actions
GET /audit-log	Full audit trail with cryptographic tamper-evidence (verify_chain())

Same logic is also exposed as **MCP tools** (`app/mcp_server.py`) and demonstrated end-to-end in an **n8n workflow** (`integrations/n8n_workflow_reconciliation.json`) — see [docs/architecture.md](docs/architecture.md).

Quickstart

python -m venv .venv && source .venv/bin/activate
pip install -r requirements-dev.txt
cp .env.example .env   # optional — leave blank to run on the local narrative fallback

uvicorn app.main:app --reload
# -> http://localhost:8000/docs for interactive Swagger UI

Try it against the sample data:

python - <<'PY'
import json, pandas as pd, requests

bank = pd.read_csv("data/sample_bank_statement.csv").to_dict(orient="records")
ledger = pd.read_csv("data/sample_internal_ledger.csv").to_dict(