How this repo works

Build, deploy, verify path

Payment Intelligence System

The L5 Payments vertical of the agentic-income ecosystem: a SIP-style payments-governance vertical plus a fail-closed TypeScript MCP server that authorizes — but never moves — money.

Status: v0.2 — hardened scaffold.

Caution

⚠️ UNAUDITED. NOT FOR LIVE FUNDS. This is a v0.2 scaffold. The MCP server verifies mandates with real Ed25519 public-key verification and keeps durable audit + replay/spend state (survives a restart), exercised by an end-to-end MCP integration test. It has still not been security-audited, it integrates no real settlement rail and no AP2 key distribution/revocation, and it must never be wired to a production payment system or live funds. There is no tool that moves money — and there never will be. Use it to model the control surface, not to settle payments.

---

Where this sits

This repo is L5 — Payments in the agentic-income layer model (agentic-ops-hub/ECOSYSTEM.md). It is the control surface that makes money safe: it answers "was this authorized, and is it within cap?" before any settlement rail runs. Capability lives below it; the swarm runtime runs above it; assurance wraps the whole stack.

L7 ASSURANCE      starlight-evals (red/blue)
L6 SWARM RUNTIME  starlight-swarm (queens + workers)
L5 PAYMENTS       payment-intelligence-system  ← YOU ARE HERE
L4 INCOME ENGINE  agentic-income-skills → agenticincome
L3 OS FAMILY      agentic-business-os / agentic-creator-os
L2 CONFIG         agentic-ops-hub
L1 CAPABILITY     agentic-creator-os (skills, agents, safety)
L0 SUBSTRATE      Starlight-Intelligence-System (SIP)

The defining invariant, inherited from the ecosystem doctrine: no autonomous money movement, ever. Agents draft, verify, and gate; humans approve capital and irreversible actions.

---

Quick map

File / dir	What
README.md	This file — purpose, placement, status, the warning banner
CLAUDE.md	Operating doctrine: fail-closed, human gate on money
AGENTS.md	Cross-tool agent card (Claude / Codex / Cursor / Gemini)
SKILL.md	The payments operating skill — when to verify, when to refuse
CANON.md	Term definitions: mandate, spend-cap, rail, settlement, SPT
MEMORY.md	Durable state, commitments, open forks — and what to record
agents/	The Payments Queen + 4 workers (markdown agent definitions)
docs/PAYMENT-PROTOCOLS.md	June 2026 state of AP2 / x402 / ACP / Visa IC / agentic-payments
mcp/	The fail-closed TypeScript MCP server (4 verify-only tools + tests)
.github/workflows/ci.yml	Build + test the MCP on Node 24

---

The control surface

Money is made safe by composition, not by trust:

1. AP2 mandate answers "was this authorized?" — a cryptographically signed proof the human approved this purchase for this amount.
2. x402 or ACP answers "how does money move?" — the settlement rail (onchain USDC, or a Shared Payment Token).

This MCP verifies the AP2 mandate and enforces spend caps before any rail settles. The four tools are all verify-only:

Tool	Job	Fail mode
verify_mandate	Reject unsigned / expired / amount-mismatched mandates	Fail closed — reject on any doubt
check_spend_cap	Per-tx / per-day / per-stream caps; single-use mandate replay guard	Over cap or replay → escalate, never auto-approve
record_audit_entry	Append-only audit log	If the log write fails, the action fails
require_human_approval	Return a pending-approval object	Never auto-approves

There is no transfer, pay, settle, or move_funds tool. None exists, by design.

---

Run the MCP

cd mcp
npm install
npm run build      # tsc → dist/
npm test           # vitest — proves a forged mandate and an over-cap spend are REJECTED

The server speaks stdio. Wire it (verify-only) to the Payments Queen — never to a worker.

---

The Payments swarm

A Payments Queen (agents/payments-queen.md) coordinates four workers and escalates capital + irreversible actions to the founder + human gate:

• mandate-verifier — proves the human authorized this charge (AP2)
• spend-cap-enforcer — enforces per-tx / day / stream caps
• settlement-auditor — writes the append-only audit entry; reconciles
• fraud-sentinel — anomaly + replay + injection detection

See agents/ and docs/PAYMENT-PROTOCOLS.md.

---

Built on SIP

Composes the Starlight Intelligence Protocol. Declines vertical canon (see CANON.md). Per SIP § Sovereignty clause.

---

Built by Frank Riemer. For builders, not consumers.