Inside the prototype: PSD2, automation and a proprietary MCP workflow
By Sergey Goncharenko
Most early-stage demos run on fixture data. Ours does not, and that distinction matters enough to write about.
The pipeline
Four stages are live end to end:
PSD2 access. Read-only account and transaction data over regulated open-banking rails. Not scraping, not screen-reading, not asking anyone for their banking password.
Cross-bank normalisation. Every bank exposes a different shape. This stage flattens them into one consistent transaction model, which is the unglamorous prerequisite for everything downstream.
An automation platform. Scheduled and event-driven workflows move data through the pipeline. We run n8n today, but the orchestration layer is deliberately swappable — using something that already exists rather than writing our own was a decision about where to spend early time.
A proprietary MCP workflow. This is the piece we consider genuinely ours. It exposes live, normalised, cross-bank financial context to an AI model through the Model Context Protocol, so the model can reason over an actual financial position rather than a description of one.
That last stage is the difference between a chatbot that discusses budgeting in the abstract and a system that can answer a question about your money because it can see it.
What that produces
A structured financial health read-out: stage, stress indicators, cashflow direction, and a written diagnosis. The public stress check on this site runs the same reasoning against an uploaded statement, so you can see the output without connecting anything.
The honest scope
The closed alpha runs on my own bank accounts. Not a cohort, not a beta group — mine.
I want to be precise about that, because "closed alpha with real banking data" is the kind of phrase that invites a generous reading. The generous reading would be wrong. One person's accounts, real data, full pipeline.
A consented friends-and-family cohort is the next step, and it is deliberately blocked on something unglamorous: writing down the consent and data-handling process properly first. Other people's bank data is not something to extend access to casually, and doing it before the paperwork exists would be exactly the wrong instinct for a company asking banks and public agencies to trust it.
What is still missing
The monitoring engine runs on schedules rather than continuously. The early-warning rules are early. And there is no accuracy figure, because nothing has been measured yet against labelled outcomes.
That measurement is the immediate priority. When there is a number, it will appear here — and not before.