Aerial view of a narrow wooden bridge connecting two lush green forested areas across calm water with reflections of the sky.
0:00
education

What’s an MCP server?

TreasurySpring
November 20, 2026
6 min

This guide explains what MCP servers are, how they let AI assistants connect directly to finance systems like accounting software and cash investment platforms, why that matters for finance and treasury teams, which risks to manage, and how to start safely, including with TreasurySpring’s own read-only MCP server.

Black and white low-angle view of a curved glass skyscraper with a wave-like, gridded facade rising against a pale sky.

AI assistants are now common in finance, but they usually aren't connected to the systems where company data is stored. To get answers about last month’s payments or this week’s maturities, you often have to export a report and paste it into a chat window.

MCP servers let AI assistants work directly with business systems so that you can skip extra steps. This guide explains what an MCP server is, how it works, why finance and treasury teams should care, what risks to watch for, and how to get started safely.

What is an MCP server?

MCP stands for Model Context Protocol. It's an open standard that explains how AI assistants connect to other software. An MCP server is the software that makes a system available through this standard. It tells the AI assistant what information the system has and what it can do, then handles each request.

Think of it like a USB-C port. Any device with the right connector works in any port, and the port's manufacturer doesn't need to know what will be plugged into it. MCP works the same way for AI: once a system has an MCP server, any compatible assistant can connect, no matter who built it.

Where MCP came from

Anthropic, the company behind Claude AI, released MCP as an open standard in November 2024. Adoption was quick, and now assistants like ChatGPT, Google Gemini, and Microsoft Copilot support it. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, a new group under the Linux Foundation co-founded with OpenAI and Block. By then, more than 10,000 MCP servers had been published.

How an MCP server works

Two pieces of software are involved. The first is the AI assistant you use, which has a built-in connector called an MCP client. The second is the MCP server, which sits in front of a business system like an accounting package or investment platform. The assistant checks which tools its connected servers offer and chooses the right one. 

The server fetches the data, applies that user’s permissions, and returns the result, which the assistant turns into a plain English answer. Depending on how the provider has built it, a server may offer read-only access or allow actions such as creating an invoice.

A worked example

A CFO at a US manufacturer is preparing the company’s next quarterly estimated federal tax payment. With MCP servers connected to the company’s accounting software and cash investment platform, they could ask which dollar investments mature before the payment is due. The assistant would check both systems, combine the results, and reply in seconds. Without MCP, the same answer means logging in to two systems, exporting two reports, and matching them in a spreadsheet. The underlying figures are identical either way, but the route is much shorter.

How an MCP server differs from an API

An application programming interface, or API, is the usual way one piece of software gets data from another. Each API is different, so connecting a system to an AI assistant used to mean writing custom code for each pairing. MCP adds a standard layer so that an assistant can find and use a system’s features without custom integration. It builds on an existing API and uses the same logins and permissions. 

The API remains the better choice for fixed, high-volume connections, such as sending prior-day bank balances to a treasury management system each morning. MCP suits the flexible, question-led work that AI assistants handle well.

Why finance teams should care about MCP

Less time spent finding information

Finance teams spend a lot of time gathering numbers that already exist in their systems. MCP lets staff ask for this information in plain English and get answers from the live system, not from last week’s export. This can free up valuable time for more value-add evaluation and decision-making.

Finance systems already support it

MCP is no longer just for developers. Intuit offers official MCP connections for QuickBooks, Oracle’s NetSuite AI Connector Service is built on MCP, and Microsoft offers an MCP server for its Dynamics 365 finance and operations apps. Treasury and cash investment platforms are adding support too, so more of the systems US finance teams rely on can now connect to an AI assistant consistently.

A foundation for automation

Right now, most finance teams use MCP to ask questions. In the future, these connections could handle scheduled reports, policy checks, and reconciliations across several systems at once. Teams that start now with read-only access and simple tasks will be able to grow with these tools as they improve. Our article on how treasurers should use AI explains where AI adds value today.

The risks to manage

Connecting AI to financial data has clear benefits, but it also creates new risks for mistakes and misuse. In spring 2026, the Cybersecurity and Infrastructure Security Agency and the National Security Agency, with cybersecurity agencies from Australia, Canada, New Zealand, and the UK, published joint guidance on the careful adoption of agentic AI. It singles out prompt injection, where concealed instructions in a file or message trick an AI model, as one of the most persistent and hardest to fix threats these systems face. The agencies recommend limiting what AI tools can access, keeping people in control of key decisions, and starting with low-risk, non-sensitive use cases. 

Starting with read-only access means the assistant can’t move money or change records. Each connection should have only the permissions it needs, every request should be logged, and any action with financial consequences should require human approval, in line with the internal controls over financial reporting that auditors already test.

Accuracy and data protection are important too. AI assistants can misread data or make mistakes in long conversations, so the source system should always be the record of truth. Companies should also check where data is processed and stored, and whether it could be used to train AI models, to meet their obligations under US state privacy laws, such as the California Consumer Privacy Act, and the confidentiality terms in their own contracts.

Questions to ask any MCP provider

Before you connect a system, check if its MCP server is read-only or can take actions, and how those actions are approved. Make sure it uses your existing logins and permissions, and that it records every request in an audit trail. Your IT and information security teams should review the connector before it’s used more widely.

How to get started with MCP

The safest way to start is with a single, low-risk use case that has a clear benefit. Querying holdings, upcoming maturities, or open invoices is a good choice, since the data already exists and the answers are easy to check. Set a short AI policy first, covering which assistants staff can use, which systems they can connect, and what data they can share.

Record the errors caught over the first few months. That evidence makes it easier to decide which connections to add next and where human review must remain. Our guide to creating an investment policy statement shows how to set the limits any AI-assisted process should respect.

MCP and cash investment at TreasurySpring

TreasurySpring has built an MCP server designed to connect portfolio data to AI assistants that support the standard. It provides read-only access to entities, current and past holdings, and the cash investments available to each client. The connection uses TreasurySpring’s authenticated API, keeps each user’s existing permissions, logs every query, and does not use client data to train AI models.

The assistant can’t place investments, move cash, or change settings through the connection. TreasurySpring intends any future ability to take actions to require human approval. Finance teams at large businesses, private funds, and other institutions can set it up after their IT team reviews the connector. Our guide to the TreasurySpring MCP server explains how.

Preparing your finance team for connected AI

An MCP server lets an AI assistant work directly with the systems your business already uses, all through one open standard. For finance teams, the main benefit is faster access to information, with automation coming as tools and controls improve. A measured approach suits most teams: start small, keep access read-only at first, and build governance as use grows.

Get in touch with TreasurySpring to find out how our MCP server can connect your cash investment data to the AI tools your team already uses.

For anyone who can’t tell the difference between their SEC and their ASC, we’ve introduced our Treasury & Cash Investment Glossary.

*TreasurySpring’s blogs and commentaries are provided for general information purposes only, and do not constitute legal, investment or other advice.

Subscribe to Insights

Explore flexible, diversified funding, built around your requirements.

We may use your contact details to follow up on your enquiry and to share information about similar products or services that may be relevant to you. You can opt out at any time. For more information, please see our privacy policy.
Subscribe
Thank you!
Your submission has been received. We will get back to you as soon as possible.
Oops! Something went wrong while submitting the form. Please try again.