Build on IXFXN
Everything the dashboard and the Assistant can do goes through the same services you can call. Same identity checks, same rules, same audit trail.
Documentation
Full documentation is being written. Until then, the overviews below cover the model and the safeguards.
REST API and Python SDK
Create and track payment requests, read balances and the ledger, and receive events. Your systems are treated like any other channel: they can request; authorised people approve.
MCP server for outside AI tools
Connect ChatGPT, Claude or your own agents through the MCP (Model Context Protocol) server. Today it is read-only: payment status, audit trails and IXFXN's reporting agents. Confirming or sending a payment always happens in IXFXN or its Slack app.
The control boundary
- Identity
- Limits
- Rules
- Compliance
- Approvals
- Routing
- Permit
- Ledger
- Audit
AI never holds the money. IXFXN decides what AI is allowed to ask for.
The payment lifecycle
Every payment moves through one state machine. An API call can't skip a state, and every move is in the audit trail.
Other outcomes: FAILED · CANCELLED · EXPIRED · SUBMISSION_UNKNOWN (never resent; the provider is queried) · PARTIALLY_SETTLED · RETURNED · REVERSED
- Permit: bound to the hash of one exact payment plan; works once; valid at most 60 seconds; compliance re-checked before it's issued.
- Idempotency: the permit is recorded as used before anything is sent, so a retry can't become a second payment.
- Separation: only the execution layer and the payment engine can reach provider connectors; AI components can't reach either. Enforced by an import check on every change.
MCP tools available today
Never exposed: signing keys, direct provider calls, ledger edits, approvals.
Python SDK, payment request with constraints planned
client.payments.create(
amount="2000000.00", currency="USD", recipient="supplier_123",
constraints={"settle_before": "16:00", "max_fee": "500", "risk": "low"},
)