Model Context Protocol is leaving developer tooling and entering production financial systems. FinBox is the latest signal, not the first.
FinBox, a credit infrastructure platform serving over 100 enterprise lenders across India, just added MCP support to its Sentinel AI decision engine. That sentence would have meant nothing 18 months ago. Now it means a protocol originally built for connecting AI models to developer tools is sitting inside credit decisioning workflows.
That is a different kind of adoption. A production system where the output is whether someone gets a loan, not a hackathon demo or a pitch-deck feature.
Model Context Protocol is no longer a developer convenience. It is becoming infrastructure for financial services.
We have been tracking MCP's evolution since the protocol first emerged from Anthropic in late 2024. The trajectory was predictable: developer tools first, then enterprise software, then regulated industries. What was not predictable was how fast that last jump would happen.
What MCP Actually Does
The simplest explanation: MCP is USB-C for AI. One standardised interface instead of custom integrations for every tool, database, and API an AI model needs to talk to.
Before MCP, connecting an AI model to a payment processor required bespoke code. Connecting it to a compliance database required different bespoke code. Every new data source meant another custom integration. That approach does not scale, and in financial services, it creates a compliance nightmare because every custom connection needs its own audit trail.
MCP changes this by providing a single open protocol that standardises how AI models discover, connect to, and interact with external systems. An AI agent running on MCP can pull transaction data, query a compliance database, and trigger a payment workflow through the same interface. The connections are structured and logged by default.
That logging part matters more than anything else in banking. We will come back to it.
Who Is Already Using It
FinBox is the headline, but the pattern started earlier.
Modern Treasury was the first fintech to implement MCP, enabling natural language interaction with payment APIs. Their implementation lets operators query ACH transfers, initiate wire payments, and monitor real-time payment flows by talking to the system instead of navigating dashboards. That is not a toy. Modern Treasury processes billions in payment volume for companies like Gusto and ClassPass.
Block has been building MCP into internal payment workflows at Square. The details are thin because Block does not talk about infrastructure work until it ships, but the signal is clear from developer job postings and open-source contributions.
Pagos brought MCP to payments data analytics, letting merchants query transaction performance data through natural language. And Fingerprint integrated MCP into its fraud prevention stack, connecting device intelligence signals to AI-driven fraud models through the protocol.
The pattern is obvious. MCP is showing up wherever AI needs to interact with financial data in real time. Credit decisioning. Payment operations. Fraud detection. Transaction analytics. These are not adjacent use cases. They are the core of what banks and fintechs do every day.
Why Banking Cares
Two words: compliance and auditability.
Financial regulators do not care what protocol your AI uses. They care whether you can prove what data the model accessed, what it did with that data, and why it made the decision it made. Every MCP connection is structured and logged. That makes it auditable in a way that ad-hoc API integrations are not.
The Coalition for Secure AI (CoSAI), an OASIS open project, published an extensive MCP security taxonomy in January 2026. That is not a trade group writing a white paper. CoSAI includes Google, Microsoft, Amazon, Intel, and Anthropic. When organisations at that scale collaborate on a security framework for a specific protocol, they are telling you what infrastructure they expect to be running in production.
The enterprise MCP market is expected to reach $1.8 billion. AWS, Azure, Google Cloud, Oracle, OpenAI, and Anthropic all shipped MCP features in their platforms. That is every major cloud and AI provider, all converging on the same protocol. 2026 is the year enterprise adoption goes from pilot to production.
Here is the thing. Banks are not adopting MCP because it is trendy. They are adopting it because the alternative is worse. The alternative is hundreds of custom integrations, each with its own security model, its own logging format, its own audit process. Regulators are already asking hard questions about AI in lending and fraud detection. A standardised, auditable connection layer is not a luxury. It is a survival mechanism.
The Gap
None of this means MCP is ready for every bank to deploy tomorrow.
MCP is a protocol, not a product. It defines how connections should work. It does not deploy them, monitor them, or secure them. That is the vendor's job, and a lot of vendors are still figuring it out.
Security is the biggest concern. Every MCP server is an endpoint. Every endpoint is an attack surface. A bank exposing its core banking system through an MCP interface needs to be very confident in its authentication, authorization, and encryption. The protocol's production readiness improvements, including better authentication, HTTP streaming, and improved error handling, are still on the 2026 roadmap. They are not shipped yet.
The gap between protocol specification and production deployment in regulated environments is where the real work happens. And that work is barely started.
There is also the human problem. Bank IT teams are not staffed to evaluate and deploy new AI protocols. Most banks are still working through their first wave of AI projects, basic chatbots and document processing. Asking them to also stand up MCP infrastructure for real-time credit decisioning is a big jump.
The banks that will move first are the ones already running AI in production for lending, fraud, or payments. They already feel the pain of managing dozens of custom integrations. For them, MCP solves a problem they have right now. For everyone else, it solves a problem they do not know they have yet.
That second group is bigger. And slower.
The FinBox announcement is not a turning point. It is a data point in a trend that has been building for months. MCP is moving from developer infrastructure to financial infrastructure, and the reasons are structural, not hype-driven. Compliance needs are real. Integration complexity is real. The demand for auditable AI connections in regulated environments is real.
The question is not whether MCP becomes standard infrastructure in financial services. The question is whether the protocol matures fast enough to meet the security and reliability standards that banking demands.
Sources
If your AI agent is already making credit decisions, who is auditing the connections it relies on?